A non-contact electrocardiogram generation method based on radar signal and enhancement strategy

Abstract

This invention discloses a non-contact ECG generation method based on radar signals and enhancement strategies. First, synchronously acquired radar heart sound signals and ECG signals are preprocessed and enhanced in multiple dimensions to construct training samples. Then, an encoder network is used to extract deep features from the radar heart sound signals step by step. Next, a multi-head self-attention mechanism and a bidirectional long short-term memory network are used to capture long-range waveform dependencies and temporal context. A high-fidelity ECG is then reconstructed via a hierarchical upsampling decoder. Finally, a four-dimensional loss function that integrates amplitude, phase, R-wave localization, and frequency domain consistency is used to optimize the model parameters. This method can reconstruct clinically diagnostic ECGs with high accuracy from non-contact radar signals. While effectively preserving key waveform morphologies such as the P wave, QRS complex, and T wave, it significantly improves waveform correlation, key feature localization accuracy, and the accuracy of long-term heart rate variability analysis.

Description

Technical Field

[0001] This invention belongs to the field of non-contact vital sign monitoring technology, specifically relating to a non-contact electrocardiogram generation method based on radar signals and enhancement strategies. Background Technology

[0002] Current electrocardiogram (ECG) monitoring relies on contact electrodes, which, while highly accurate, suffer from issues such as discomfort, skin allergies, and prolonged wear, limiting their applicability in scenarios like sleep monitoring, burn patient monitoring, and emotion computing. In recent years, radar-based non-contact monitoring technology has gained widespread attention due to its advantages such as being contactless, privacy-preserving, and able to penetrate clothing. This technology extracts vital signs related to cardiac mechanical activity by emitting electromagnetic waves and receiving echoes reflected from subtle movements of the chest cavity.

[0003] However, reconstructing clinically diagnostic ECGs from radar signals still faces multiple challenges: the micro-motion signals received by radar are extremely weak and easily affected by environmental noise, random human movement, and system noise; there is a complex nonlinear mapping relationship between cardiac mechanical activity and electrophysiological activity, and existing signal processing methods are unable to reconstruct complete ECG waveforms containing key features such as P waves, QRS complexes, and T waves; in addition, the scarcity of precisely synchronized paired data between radar and ECG signals restricts the training of data-driven models.

[0004] Previous studies have attempted to utilize deep learning for radar ECG reconstruction. The paper "Chen, Jinbo, et al. 'Contactless Electrocardiogram Monitoring With Millimeter Wave Radar. MobileComputing, IEEE Transactions on 23.1(2024):16'" proposes a non-contact ECG monitoring method based on millimeter-wave radar, achieving end-to-end reconstruction by fusing CNN and Transformer. The paper "Yamamoto, Kohei, R. Hiromatsu, and T. Ohtsuki. 'ECG Signal Reconstruction via Doppler Sensor by Hybrid Deep Learning Model with CNN and LSTM.' IEEE Access PP.99(2020):1-1'" proposes a hybrid deep learning model for reconstructing ECG signals using a Doppler sensor, improving temporal consistency by combining CNN and LSTM networks. However, existing methods still have shortcomings in terms of waveform detail restoration, noise robustness, and key feature localization accuracy. Moreover, most of them rely on body surface displacement signals that have millisecond-level delays with ECG activity, failing to fully explore radar heart sound signals that are more synchronous and direct with ECG events in terms of time sequence.

[0005] Therefore, there is an urgent need for a method that can reconstruct clinically usable ECG waveforms from radar signals with high accuracy and robustness, in order to promote the practical application of non-contact ECG monitoring. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a non-contact electrocardiogram (ECG) generation method based on radar signals and enhancement strategies (Radar-based, Augmentation-enhanced Generation of Electrocardiogram, RAE-ECG), which reconstructs a high-fidelity ECG with clinical diagnostic value from thoracic micromotion signals acquired by radar.

[0007] The technical solution adopted in this invention is: a non-contact ECG generation method based on radar signals and enhancement strategies, the specific steps of which are as follows:

[0008] S1. Perform data collection and preprocessing to construct a paired dataset;

[0009] First, the synchronously acquired radar signals and corresponding electrocardiogram signals are used as training and testing data for the model.

[0010] Among them, the radar signal is the I / Q signal of the radar acquired by the continuous wave radar system; the electrocardiogram signal is the synchronous contact ECG record with clinical reference value, that is, the reference ECG signal.

[0011] Then, the radar I / Q signals are preprocessed, that is, elliptic correction is achieved by fitting the I / Q trajectory to the ideal unit circle. The generalized quadratic curve model expression is as follows:

[0012] ;

[0013] in, and These represent the in-phase component and quadrature component of the radar I / Q signal, respectively. , , , , , All coefficients are real numbers, and , , Determines the shape of the trajectory. , Reflects the DC bias of the signal. This is a constant term.

[0014] Reconstruct the data matrix , its first Behavior . Indicates the first The values ​​of the in-phase components at each data sampling point correspond to the I component in the raw I / Q signal received by the radar. Indicates the first The values ​​of the orthogonal components of each data sampling point correspond to the Q component in the original radar I / Q signal.

[0015] Then solve the matrix equation. ,in The coefficient vector corresponding to the minimum singular value is obtained by using singular value decomposition (SVD). The least squares estimation of the model parameters is achieved; then, based on the solved coefficients, an inverse correction transformation is constructed to correct the distorted I / Q trajectory and map it to the vicinity of the ideal unit circle around the origin.

[0016] in, This represents the matrix transpose operation.

[0017] Then, the radar heart sound signal is extracted from the corrected phase signal, the heart sound is separated by a 0–80 Hz bandpass filter, and aligned with the preprocessed reference ECG signal on the time axis.

[0018] Among them, the ECG signal is preprocessed, that is, a wavelet threshold denoising algorithm is used to suppress power frequency interference, electromyographic noise and baseline drift, improve the signal-to-noise ratio and retain the key waveform shape.

[0019] The key waveforms include: P-wave, QRS complex, and T-wave.

[0020] Finally, the preprocessed long-time-series signal, namely the preprocessed radar heart sound signal and the reference ECG signal, is segmented into fixed segments of a predetermined length, ensuring that each segment includes 2–3 complete cardiac cycles. The preprocessed radar heart sound signal and the reference ECG signal are then time-aligned and amplitude-normalized to form paired "radar signal – ECG signal" samples. The final dataset contains N paired samples, divided into training, validation, and test sets according to a predetermined ratio.

[0021] Before the training phase begins, multi-dimensional data augmentation is performed on the training set samples, including: temporal pruning and interpolation, time axis scaling, amplitude scaling, frequency domain selective filtering, and adaptive Gaussian noise injection.

[0022] S2. Based on step S1, construct the RAE-ECG generation model;

[0023] A deep sequence generation model, namely the RAE-ECG generation model, is constructed. This model consists of three cascaded modules: a feature extraction encoder, a temporal modeling module, and a sequence generation decoder.

[0024] S3. Based on the RAE-ECG generation model constructed in step S2, input the paired dataset constructed in step S1, train and test the model, and output the reconstructed ECG.

[0025] First, the training set is input into the RAE-ECG generation model for model training. That is, the radar heart sound signal after front-end preprocessing is used as input, and the real electrocardiogram signal collected at the same time is used as the supervision target. The model is optimized in an end-to-end manner.

[0026] The training objective of the RAE-ECG generative model employs a four-dimensional composite loss function. The driver, this function, is driven by weighting coefficients. , , , Coordinate the four sub-items to simultaneously optimize the time-domain waveform similarity of the reconstructed signal. Phase consistency Accuracy of localization of key diagnostic features and frequency domain energy distribution The expression is as follows:

[0027] ;

[0028] Among them, the key diagnostic features include: P wave, R wave and T wave.

[0029] Then the RAE-ECG generation model is trained using the mini-batch gradient descent strategy, and the specific settings are as follows:

[0030] The optimizer selects AdamW, and the weight decay coefficient λ > 0 is set; the gradient accumulation strategy is adopted, and the weights are updated every k steps; the learning rate scheduling applies the OneCycle strategy, linearly rising to the peak η_max > 0 within the first r% of the training steps, where 0 < r < 100, and decaying through the cosine annealing function in the remaining steps.

[0031] Among them, the upper limit of the total number of epochs in the training process is set to T_max, and a dynamic early stopping mechanism is introduced, that is, the validation set is input for verification. After w epochs of training start, if the relative improvement amplitude of the validation set loss for p consecutive epochs does not exceed the threshold δ > 0, the training is automatically terminated.

[0032] After the RAE-ECG generation training is completed, the performance of the model is evaluated using an independent test set. The radar signal segments of the test set are input into the trained model, that is, input into the feature extraction encoder, the time series modeling module, and the sequence generation decoder in sequence, and the corresponding electrocardiogram waveform segments are automatically reconstructed. And to evaluate the practicality of the model in long-term heart rhythm analysis, an overlapping segmentation and weighted merging strategy is adopted to reconstruct the continuous electrocardiogram signal from the segments, that is, the original long sequence radar signal is segmented and input with 50% overlap. After independent reconstruction of each segment, the overlapping regions are weighted and averaged and fused using a Hanning window, and finally a continuous and complete high-fidelity electrocardiogram signal is generated.

[0033] S4. Based on the reconstructed ECG output in step S3, a systematic evaluation is carried out according to the evaluation indicators to obtain a clinically available ECG, and the non-contact to clinical-grade electrocardiogram generation is completed;

[0034] The evaluation indicators include: waveform correlation, key wave positioning, spectrum consistency, and HRV analysis.

[0035] First, evaluate the overall waveform similarity of the reconstructed ECG, then ensure the spectrum characteristics of the reconstructed signal and the real ECG are consistent within the clinical core frequency band through time-frequency analysis, then evaluate the positioning accuracy of the model for key diagnostic features, and confirm its time synchronization and the integrity of the morphological structure; finally, perform heart rate variability HRV analysis on the reconstructed continuous ECG signal.

[0036] Among them, the key diagnostic features include: P wave, R wave and T wave.

[0037] After multi-dimensional verification by evaluation indicators, the model stably converts the input non-contact radar signal into a standard ECG signal with clear waveform characteristics, complete time and frequency information, and usable for heart rhythm analysis, realizing the generation of ECG signals from non-contact to clinical level.

[0038] Furthermore, in step S2, the RAE-ECG generation model is specifically as follows:

[0039] A1. Feature Extraction Encoder: Employs a hierarchical, multi-stage convolutional architecture, which continuously downsamples and transforms features to progressively refine the original waveform into a feature representation rich in multi-scale contextual information.

[0040] The feature extraction encoder receives the preprocessed radar heart sound signal sequence as input. The encoding process consists of four stages: The first stage uses a one-dimensional convolutional kernel of length 15 for initial feature mapping and maintains the temporal length through mirror padding, increasing the number of channels to 64; The second and third stages each consist of two cascaded residual blocks, gradually increasing the number of channels to 128 and using max pooling with strides of 4 and 5 to compress the temporal dimension, thereby expanding the receptive field and suppressing noise; The fourth stage stacks three residual blocks, where the convolutional kernel size decreases progressively with sizes of 7, 5, and 3, and the number of channels is fixed at 256. Multi-scale feature fusion is achieved through the continuous stacking of small convolutional kernels; Finally, the feature extraction encoder outputs a high-level feature sequence with compressed temporal dimension and expanded channel dimension.

[0041] Each residual block includes batch normalization and exponential linear unit activation, and mitigates gradient vanishing through skip connections, supporting deeper network stacking. The structural expression for each residual block is as follows:

[0042] ;

[0043] in, and These represent the input and output tensors of the residual block, respectively. This represents the weight set { Parameterized main path nonlinear transformation; Indicating bypass The weight matrix of the convolution is used to weight the input. Perform the necessary linear projection to match the dimensions; the operator '+' indicates element-wise addition.

[0044] A2, Timing Modeling Module;

[0045] The temporal modeling module adopts an architecture based on self-attention mechanism and bidirectional long short-term memory network to model the long-term dependencies within the feature sequence and focuses on key temporal segments related to ECG features.

[0046] The temporal modeling module receives the feature sequence encoded by the feature extraction encoder, including: two multi-head attention modules and a bidirectional long short-term memory network (Bi-LSTM) module.

[0047] First, the feature sequence encoded by the feature extraction encoder is input into the multi-head attention module. That is, the multi-head attention mechanism captures the long-range dependencies of the sequence, splitting the features into 8 heads, and each head independently calculates attention, as shown in the following expression:

[0048] ;

[0049] in, This represents the dimension of the key vector, used as a scaling factor to stabilize the gradient; Indicates the attention head index; This represents the normalization function along the last dimension. , , Let these represent the generated query matrix, key matrix, and value matrix, respectively, and satisfy the following expression:

[0050]

[0051] in, Represents the input feature matrix; They represent the first The corresponding trainable query, key, and value projection weight matrix for each head;

[0052] The outputs of all attention heads in each multi-head attention module are concatenated and linearly transformed to obtain the final output of the multi-head attention module.

[0053] Then, the Bi-LSTM (Bi-Long Short-Term Memory) network module receives the sequence after attention weighting by all multi-head attention modules, and uses bi-directional recursive connections to accurately model the quasi-periodic patterns and long-range temporal dependencies within the heartbeat cycle, as shown in the following expression:

[0054] ;

[0055] ;

[0056] ;

[0057] in, Indicates the time step index of the sequence; Indicates time step The input feature vector; and These represent the forward and backward LSTM units at time steps, respectively. The hidden state; and These represent the hidden states at the previous time step in the corresponding direction; Represents a computation function that includes gating mechanisms; operators This represents a vector concatenation operation. LSTM units regulate the information flow through forget gates, input gates, and output gates.

[0058] A3, Sequence Generation Decoder;

[0059] The sequence generation decoder receives the high-level feature representation after attention weighting and temporal modeling in the temporal modeling module. The decoder employs a cascaded architecture of hierarchical upsampling and residual refinement.

[0060] First, the feature tensor is processed by a local connection layer to smooth the feature distribution and suppress high-frequency noise. Then, it enters the first stage of upsampling, where the sequence length is extended to 1250 points based on the principle of spectrum shifting, and a double residual block is embedded to filter and optimize the extended signal. Then, it enters the second stage of upsampling, where the signal is stretched to a full length of 5000 points, and a physiological feature guidance mechanism is introduced. That is, by synergistically using multi-scale convolution kernels, the reconstruction process conforms to the physiological laws of ECG signals, and finally outputs a standardized ECG signal sequence that is clinically usable in both the temporal morphology and frequency distribution.

[0061] Furthermore, in step S3, the four-dimensional composite loss function is specifically as follows:

[0062] Time-domain waveform similarity of reconstructed signals That is, amplitude loss, which is addressed by introducing the attention weight matrix of the ECG R wave. The weighted sum is expressed as follows:

[0063] ;

[0064] in, Indicates the amplitude loss value; This represents the expected value, and the batch average is taken in actual training. and These represent the predicted waveform of the non-contact ECG signal and the actual contact ECG reference waveform, respectively. This represents absolute value operations.

[0065] Phase consistency Phase loss, measured by first-order difference cosine similarity, is expressed as follows:

[0066] ;

[0067] in, This represents the phase loss value; Indicates the total length of the electrocardiogram (ECG) signal sequence; Indicates the time step index; The first-order difference operator is defined as follows: It is used to capture instantaneous changes in signals; and These represent the time steps. The predicted ECG signal value and the actual ECG signal value; This represents the function for calculating cosine similarity.

[0068] R-wave positioning accuracy That is, the R-wave focusing loss, and the L1 loss calculated in the critical region, as shown in the following expression:

[0069] ;

[0070] in, denoted as R-wave focus loss value; A represents the ECG R-wave attention weight matrix.

[0071] Frequency domain energy distribution That is, the frequency domain loss, combined with the short-time Fourier transform amplitude and phase error, is expressed as follows:

[0072] ;

[0073]

[0074]

[0075] in, This represents the total frequency domain loss value; and These represent the amplitude spectrum loss and the phase spectrum loss, respectively. Represents the short-time Fourier transform; Indicated in frequency index and time frame index Amplitude spectrum value at; and Indicates the corresponding phase value; and These represent the total number of frequency domain cells and the total number of time domain frames, respectively. and Indicates the summation index; This represents summing over all time-frequency units; This represents the normalization factor for calculating the average. 0.5 represents the sine function; 0.5 is the weighting coefficient for phase loss.

[0076] The beneficial effects of this invention are as follows: The method of this invention first preprocesses and enhances synchronously acquired radar heart sound signals and electrocardiogram (ECG) signals in multiple dimensions to construct training samples. Then, an encoder network is used to extract deep features from the radar heart sound signals step by step. Next, a multi-head self-attention mechanism and a bidirectional long short-term memory network are used to capture long-range waveform dependencies and temporal context. A high-fidelity ECG is then reconstructed via a hierarchical upsampling decoder. Finally, a four-dimensional loss function that integrates amplitude, phase, R-wave localization, and frequency domain consistency is used to optimize the model parameters. This method can reconstruct clinically diagnostic ECGs with high precision from non-contact radar signals. While effectively preserving key waveform morphologies such as the P wave, QRS complex, and T wave, it significantly improves waveform correlation, key feature localization accuracy, and the accuracy of long-term heart rate variability analysis. It is suitable for applications that do not require electrode contact, such as remote health monitoring, long-term home monitoring, sleep analysis, and driving safety warnings. Attached Figure Description

[0077] Figure 1 This is a flowchart of a non-contact ECG generation method based on radar signals and enhancement strategies according to the present invention.

[0078] Figure 2 This is a schematic diagram of radar I / Q signal elliptic correction in an embodiment of the present invention.

[0079] Figure 3 This is a comparison diagram of the timing alignment between radar heart sounds and electrocardiogram signals in an embodiment of the present invention.

[0080] Figure 4 This is a network structure diagram of the RAE-ECG model in an embodiment of the present invention.

[0081] Figure 5 This is an R-wave attention weight distribution diagram in an embodiment of the present invention.

[0082] Figure 6 This is a comparison diagram of the full waveform, partial waveform, and spectrum of the reconstructed ECG and the real ECG in an embodiment of the present invention.

[0083] Figure 7 This is a statistical chart showing the distribution of PCC and NRMSE indicators in an embodiment of the present invention.

[0084] Figure 8 This is a box plot of the P / R / T wave positioning error in an embodiment of the present invention.

[0085] Figure 9 This is a comparison image of long-sequence continuous ECG reconstruction in an embodiment of the present invention. Detailed Implementation

[0086] The method of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0087] like Figure 1 The flowchart of a non-contact ECG generation method based on radar signals and enhancement strategies of the present invention is shown below. The specific steps are as follows:

[0088] S1. Perform data collection and preprocessing to construct a paired dataset;

[0089] To generate an electrocardiogram (ECG) from non-contact radar signals, it is first necessary to acquire synchronously collected radar signals and corresponding ECG signals as training and testing data for the model.

[0090] Among them, the radar signal can reflect the micro-movements of the chest cavity caused by the mechanical activity of the heart, and the I / Q signal of the radar is collected by a continuous wave radar system; the electrocardiogram signal is a synchronous contact ECG record with clinical reference value, that is, the reference ECG signal.

[0091] like Figure 2 As shown, considering the non-ideal hardware characteristics of radar systems such as amplitude imbalance, phase shift, or DC bias, preprocessing of the radar I / Q signals is required. This involves fitting the I / Q trajectories to an ideal unit circle to achieve elliptic correction (to compensate for amplitude imbalance, quadrature phase deviation, and DC bias that may be introduced by the non-ideal characteristics of the RF front end). The generalized quadratic curve model expression is as follows:

[0092] ;

[0093] in, and These represent the in-phase component and quadrature component of the radar I / Q signal, respectively. , , , , , All coefficients are real numbers, and , , Determines the shape of the trajectory. , Reflects the DC bias of the signal. These are constant terms. By estimating these parameters, a corrective transformation from a non-ideal ellipse to an ideal unit circle can be constructed.

[0094] Reconstruct the data matrix , its first Behavior . Indicates the first The values ​​of the in-phase components at each data sampling point. In radar signal processing, this typically corresponds to the I component in the raw I / Q signal received by the radar. Indicates the first The values ​​of the orthogonal components of each data sampling point correspond to the Q component in the original radar I / Q signal.

[0095] Then solve the matrix equation. ,in The coefficient vector corresponding to the minimum singular value is obtained by using singular value decomposition (SVD). The least squares estimation of model parameters is achieved; then, based on the solved coefficients, an inverse correction transformation is constructed to correct the distorted I / Q trajectory and map it to the vicinity of the ideal unit circle around the origin, thereby compensating for the system error introduced by the hardware at the algorithm level.

[0096] in, This represents the matrix transpose operation.

[0097] like Figure 3 As shown, the radar heart sound signal is then extracted from the corrected phase signal, separated by a 0–80 Hz bandpass filter, and aligned with the preprocessed reference ECG signal on the time axis.

[0098] Among them, the ECG signal is preprocessed, that is, a wavelet threshold denoising algorithm is used to suppress power frequency interference, electromyographic noise and baseline drift, improve the signal-to-noise ratio and retain the key waveform shape.

[0099] The key waveforms include: P-wave, QRS complex, and T-wave.

[0100] Finally, the preprocessed long-time-series signal, namely the preprocessed radar heart sound signal and the reference ECG signal, is segmented into segments of a preset fixed length (e.g., 5000 sampling points, corresponding to approximately 2.5 seconds), ensuring that each segment includes 2–3 complete cardiac cycles. The preprocessed radar heart sound signal and the reference ECG signal are then time-aligned and amplitude-normalized to form "radar signal – ECG signal" paired samples. The final dataset contains N paired samples (N is a positive integer determined based on actual conditions), divided into training, validation, and test sets according to a preset ratio. In this embodiment, the training set accounts for 80%, the validation set for 10%, and the test set for 10%, used for subsequent model training, optimization, and evaluation.

[0101] To improve the model's robustness to noisy environments and individual differences, multi-dimensional data augmentation is performed on the training set samples before the training phase begins. This includes temporal pruning and interpolation, time axis scaling, amplitude scaling, frequency-domain selective filtering, and adaptive Gaussian noise injection. This augmentation process is applied only during model training and does not alter the original distribution of the test and validation sets.

[0102] S2. Based on step S1, construct the RAE-ECG generation model;

[0103] like Figure 4 As shown, in order to achieve end-to-end mapping from radar cardiac signals to standard electrocardiogram waveforms, a deep sequence generation model, namely the RAE-ECG generation model, is constructed. This model includes three cascaded modules: a feature extraction encoder, a time-series modeling module, and a sequence generation decoder.

[0104] S3. Based on the RAE-ECG generation model constructed in step S2, input the paired dataset constructed in step S1, train and test the model, and output the reconstructed ECG.

[0105] First, the training set is input into the RAE-ECG generation model for model training. That is, the radar heart sound signal after front-end preprocessing is used as input, and the real electrocardiogram signal collected at the same time is used as the supervision target. The model is optimized in an end-to-end manner.

[0106] The training objective of the RAE-ECG generative model employs a four-dimensional composite loss function. The driver, this function, is driven by weighting coefficients. , , , Coordinate the four sub-items to simultaneously optimize the time-domain waveform similarity of the reconstructed signal. Phase consistency Accuracy of localization of key diagnostic features and frequency domain energy distribution The expression is as follows:

[0107] ;

[0108] The key diagnostic features include: P wave, R wave and T wave.

[0109] The RAE-ECG generative model is then trained using a mini-batch gradient descent strategy, with specific settings including:

[0110] The optimizer used is AdamW, and a weight decay coefficient λ (λ > 0, usually a small value; in this embodiment, a value of 0.0001 is used for reference based on actual conditions) is set to enhance regularization. A gradient accumulation strategy is adopted, updating the weights every k steps (k is a positive integer, set according to hardware conditions; in this embodiment, a value of 4 is used for reference based on actual conditions) to effectively expand the effective batch size under limited computing resources. The learning rate scheduling uses the OneCycle strategy, linearly increasing to the peak value η_max (η_max > 0, usually a small value greater than λ; in this embodiment, a value of 0.0003 is used for reference based on actual conditions) within the first r% (0 < r < 100, set based on experience; a value of 30 is used for reference based on actual conditions) of training steps, and decaying through the cosine annealing function in the remaining steps, aiming to accelerate convergence and improve model accuracy.

[0111] The total number of training rounds is set to an upper limit of T_max (T_max is a positive integer, representing the maximum number of training rounds, and is determined based on actual conditions). A dynamic early stopping mechanism is introduced, where a validation set is input for validation. After training starts for w rounds (w is a positive integer used for warm-up, and in this embodiment, it is taken as 25 based on actual conditions), if the relative improvement of the validation set loss for p consecutive rounds (p is a positive integer, and in this embodiment, it is taken as 35 based on actual conditions) does not exceed the threshold δ (δ > 0, representing the improvement rate, and in this embodiment, it is set to 0.05%), then training is automatically terminated to suppress overfitting and improve computational efficiency.

[0112] After the RAE-ECG generation training is completed, the model performance is evaluated using an independent test set. Radar signal segments from the test set are input into the trained model, sequentially into the feature extraction encoder, temporal modeling module, and sequence generation decoder, automatically reconstructing the corresponding ECG waveform segments. To evaluate the model's practicality in long-term heart rhythm analysis, an overlapping segmentation and weighted merging strategy is used to reconstruct continuous ECG signals from the segments. Specifically, the original long-sequence radar signal is segmented with 50% overlap, and after independent reconstruction of each segment, the overlapping areas are weighted and averaged using a Hanning window to ultimately generate a continuous and complete high-fidelity ECG signal (reconstructed ECG) for subsequent waveform index analysis and heart rate variability (HRV) assessment.

[0113] S4. Based on the reconstructed ECG output from step S3, a systematic evaluation is performed according to the evaluation indicators to obtain a clinically usable ECG, thus completing the non-contact to clinical-grade ECG generation.

[0114] The evaluation metrics include: waveform correlation, key wave localization, spectral consistency, and HRV analysis.

[0115] First, the overall waveform similarity of the reconstructed ECG was assessed to verify the fidelity of the reconstructed signal in terms of morphology and amplitude. Then, time-frequency analysis was used to ensure that the spectral characteristics of the reconstructed signal and the real ECG were consistent in the clinical core frequency band. Next, the model's localization accuracy of key diagnostic features was evaluated to confirm its time synchronization and the integrity of its morphological structure. Finally, heart rate variability (HRV) analysis was performed on the reconstructed continuous ECG signal to verify its clinical usability in long-term rhythm analysis and autonomic nervous function assessment.

[0116] The key diagnostic features include: P wave, R wave and T wave.

[0117] After multi-dimensional verification by evaluation indicators, the model can stably convert the input non-contact radar signal into a standard ECG signal with clear waveform characteristics, complete time and frequency information, and usable for heart rhythm analysis, thus achieving the complete technical goal of non-contact to clinical-grade ECG generation.

[0118] In this embodiment, the RAE-ECG generation model in step S2 is as follows:

[0119] A1. Feature Extraction Encoder: Employs a hierarchical, multi-stage convolutional architecture, which continuously downsamples and transforms features to progressively refine the original waveform into a feature representation rich in multi-scale contextual information.

[0120] The feature extraction encoder receives the preprocessed radar heart sound signal sequence as input. The encoding process consists of four stages: The first stage uses a one-dimensional convolutional kernel of length 15 for initial feature mapping and maintains the temporal length through mirror padding, increasing the number of channels to 64; The second and third stages each consist of two cascaded residual blocks, gradually increasing the number of channels to 128 and using max pooling with strides of 4 and 5 to compress the temporal dimension, thereby expanding the receptive field and suppressing noise; The fourth stage stacks three residual blocks, where the convolutional kernel size decreases progressively with sizes of 7, 5, and 3, and the number of channels is fixed at 256. Multi-scale feature fusion is achieved through the continuous stacking of small convolutional kernels; Finally, the feature extraction encoder outputs a high-level feature sequence with compressed temporal dimension and expanded channel dimension, providing the subsequent temporal modeling module with a feature representation that can effectively capture local transient features and macroscopic rhythmic patterns within the heartbeat cycle.

[0121] Each residual block includes batch normalization and exponential linear unit activation, and mitigates gradient vanishing through skip connections, supporting deeper network stacking. The structural expression for each residual block is as follows:

[0122] ;

[0123] in, and These represent the input and output tensors of the residual block, respectively. This represents the weight set { Parameterized main path nonlinear transformation; Indicating bypass The weight matrix of the convolution is used to weight the input. Perform the necessary linear projection to match the dimensions; the operator '+' indicates element-wise addition.

[0124] A2, Timing Modeling Module;

[0125] The temporal modeling module adopts an architecture based on self-attention mechanism and bidirectional long short-term memory network to model the long-term dependencies within the feature sequence and focus on key temporal segments related to ECG features, thereby capturing quasi-periodic patterns and morphological details within the heartbeat cycle and inputting them into the temporal modeling module.

[0126] The temporal modeling module receives the feature sequence encoded by the feature extraction encoder, including: two multi-head attention modules and a bidirectional long short-term memory network (Bi-LSTM) module.

[0127] First, the feature sequence encoded by the feature extraction encoder is input into the multi-head attention module. That is, the multi-head attention mechanism captures the long-range dependencies of the sequence, splitting the features into 8 heads, and each head independently calculates attention, as shown in the following expression:

[0128] ;

[0129] in, This represents the dimension of the key vector, used as a scaling factor to stabilize the gradient; Indicates the attention head index; This represents the normalization function along the last dimension. , , Let these represent the generated query matrix, key matrix, and value matrix, respectively, and satisfy the following expression:

[0130]

[0131] in, Represents the input feature matrix; They represent the first The corresponding trainable query, key, and value projection weight matrix for each head;

[0132] The outputs of all attention heads in each multi-head attention module are concatenated and linearly transformed to obtain the final output of the multi-head attention module.

[0133] Then, the Bi-LSTM (Bi-Long Short-Term Memory) network module receives the sequence after attention weighting by all multi-head attention modules, and uses bi-directional recursive connections to accurately model the quasi-periodic patterns and long-range temporal dependencies within the heartbeat cycle, as shown in the following expression:

[0134] ;

[0135] ;

[0136] ;

[0137] in, Indicates the time step index of the sequence; Indicates time step The input feature vector; and These represent the forward and backward LSTM units at time steps, respectively. The hidden state; and These represent the hidden states at the previous time step in the corresponding direction; Represents a computation function that includes gating mechanisms; operators This represents a vector concatenation operation. LSTM units regulate the information flow through forget gates, input gates, and output gates to enhance the modeling ability of the periodicity and waveform temporal structure of ECG signals.

[0138] A3, Sequence Generation Decoder;

[0139] The sequence generation decoder receives the high-level feature representation after attention weighting and temporal modeling in the temporal modeling module. The decoder employs a cascaded architecture of hierarchical upsampling and residual refinement.

[0140] First, the feature tensor is processed by a local connection layer to smooth the feature distribution and suppress high-frequency noise. Then, in the first stage of upsampling, the sequence length is extended to 1250 points based on the principle of spectrum shifting, and a dual residual block is embedded to filter and optimize the extended signal. This design significantly reduces the reconstruction error of transient features such as the QRS complex. Next, in the second stage of upsampling, the signal is stretched to a full length of 5000 points, and a physiological feature guidance mechanism is introduced. This involves the coordinated use of multi-scale convolutional kernels to ensure that the reconstruction process conforms to the physiological laws of ECG signals—larger convolutional kernels are responsible for reconstructing low-frequency components and maintaining baseline stability, medium-sized convolutional kernels enhance the dynamics of waveforms such as the P wave and T wave, while smaller convolutional kernels specifically capture microvolt-level high-frequency details in the QRS complex. This design allows the decoder to fuse features of different scales layer by layer, ultimately outputting a standardized ECG signal sequence that is clinically usable in both the temporal morphology and frequency distribution.

[0141] In this embodiment, the four-dimensional composite loss function in step S3 is as follows:

[0142] Time-domain waveform similarity of reconstructed signals That is, amplitude loss, which is addressed by introducing the attention weight matrix of the ECG R wave. The weighted sum is expressed as follows:

[0143] ;

[0144] in, Indicates the amplitude loss value; This represents the expected value, and the batch average is taken in actual training. and These represent the predicted waveform of the non-contact ECG signal and the actual contact ECG reference waveform, respectively. This represents absolute value operation; this loss function makes the model focus on overall waveform fitting during training.

[0145] Phase consistency Phase loss, measured by first-order difference cosine similarity, is expressed as follows:

[0146] ;

[0147] in, This represents the phase loss value; Indicates the total length of the electrocardiogram (ECG) signal sequence; Indicates the time step index; The first-order difference operator is defined as follows: It is used to capture instantaneous changes in signals; and These represent the time steps. The predicted ECG signal value and the actual ECG signal value; This represents the function for calculating cosine similarity.

[0148] This loss function effectively ensures the correct position of key features such as R-waves and T-waves on the time axis in the generated waveform by penalizing the difference in the trend of change between the predicted signal and the real signal.

[0149] R-wave positioning accuracy That is, the R-wave focusing loss, and the L1 loss calculated in the critical region, as shown in the following expression:

[0150] ;

[0151] in, Represents the R-wave focus loss value; A represents the ECG R-wave attention weight matrix, which is generated based on the R-wave position markings in the actual ECG, such as... Figure 5 As shown, its matrix elements have larger values ​​at the time positions corresponding to the R-wave, while the values ​​are smaller at other positions, thus highlighting the accuracy of the R-wave amplitude in the loss calculation.

[0152] Frequency domain energy distribution That is, the frequency domain loss, combined with the short-time Fourier transform amplitude and phase error, is expressed as follows:

[0153] ;

[0154]

[0155]

[0156] in, This represents the total frequency domain loss value; and These represent the amplitude spectrum loss and the phase spectrum loss, respectively. Represents the short-time Fourier transform; Indicated in frequency index and time frame index Amplitude spectrum value at; and Indicates the corresponding phase value; and These represent the total number of frequency domain cells and the total number of time domain frames, respectively. and Indicates the summation index; This represents summing over all time-frequency units; This represents the normalization factor for calculating the average. represents the sine function; 0.5 is the weighting coefficient for phase loss. This loss function ensures that the distribution characteristics of the generated signal in the joint time-frequency domain are consistent with those of the target signal.

[0157] To verify the effectiveness of the method of this invention, this embodiment further conducted experimental verification. First, synchronously acquired radar signals and electrocardiogram (ECG) signals were prepared as the basis for model training and testing. Before implementation, synchronous signal data containing samples from multiple subjects were acquired, wherein the radar signal sampling rate should preferably be no less than 200Hz to capture the details of chest cavity micro-movements; the synchronously acquired contact ECG signals should have clinical reference value, and a precise time synchronization relationship between the two should be ensured.

[0158] In this embodiment, the model input is a single-channel 5000-point radar heart sound segment, and the output is the corresponding ECG waveform. The AdamW optimizer is used with an initial learning rate of 0.0003, employing a OneCycle scheduling strategy. The batch size is 32, the maximum training epochs are 500, and an early stopping mechanism is implemented (the model stops if the loss decreases by less than 0.05% after 35 consecutive validation epochs). The loss function weights are set to... The training environment consisted of an Intel Core i7-13650HX processor and an NVIDIA RTX 4070 GPU, using mixed precision training acceleration.

[0159] The performance of the RAE-ECG model based on the method of this invention has been systematically verified through multiple sets of experiments. For example... Figure 6As shown, the ECG waveform reconstructed by the model (red dashed line) highly overlaps with the real signal (blue solid line) within a continuous 2.5-second range. Key morphological features such as the P wave, QRS complex, and T wave are accurately restored. In particular, no significant distortion or time shift is observed in critical diagnostic regions such as the R wave peak and ST segment, indicating that the model possesses the ability to accurately capture dynamic changes in electrocardiograms. To quantify overall similarity, the Pearson correlation coefficient (PCC) and normalized root mean square error (NRMSE) distributions for all samples in the test set are shown below. Figure 7 As shown, the PCC values ​​are concentrated above 0.9, with an average of 0.954; while the NRMSE is mainly below 0.075, with an average of only 0.036. These two indicators together confirm that the reconstructed waveform has extremely high fidelity and extremely low amplitude error.

[0160] Regarding the positioning accuracy of key waveform features, Figure 8 The box plots show that the R wave, due to its steep slope and high amplitude, is the easiest to capture, with a median localization error of only 2.47 milliseconds. The median localization errors for the T wave and P wave are 3.43 milliseconds and 4.61 milliseconds, respectively. Although slightly higher than the R wave, these errors are still well within the clinically acceptable tolerance range, consistent with the physiological characteristics of each waveform and their susceptibility to noise. The model's performance in long sequence reconstruction is as follows... Figure 9 As shown, the 10-20 second continuous ECG signal reconstructed by the overlapping segment weighted merging strategy not only maintains the waveform accuracy within the segment, but also shows no obvious distortion at the splicing points between segments, proving that it is suitable for continuous monitoring scenarios.

[0161] Heart rate variability (HRV) analysis was further performed on the reconstructed long-sequence signal, and the results are shown in Table 1.

[0162] Table 1

[0163]

[0164] As shown in Table 1, the core HRV indices calculated from the reconstructed signal—SDNN, RMSSD, and pNN50—are highly close to the results of the actual signal, with relative errors of 0.80%, 1.34%, and 6.86%, respectively. The errors of SDNN (reflecting overall variability) and RMSSD (reflecting short-term variability) are both within 1.5%, indicating that the model excellently preserves the rhythmic information of the heart rate. Although the error of pNN50 is slightly larger, it is still below the clinically acceptable threshold of 7%. Based on waveform comparison, feature localization, and HRV analysis, the method of this invention can reconstruct high-quality electrocardiograms with clinical diagnostic value in terms of morphology, temporal sequence, and physiological rhythm from non-contact radar signals, providing a reliable technical foundation for non-intrusive health monitoring.

[0165] In summary, the method of this invention achieves high-precision reconstruction from radar heart sound signals to clinical-grade electrocardiograms through a triple enhancement design. Compared with existing technologies, the method of this invention has the following advantages: 1. Using radar heart sounds as input, it more directly correlates with cardiac mechanical and electrical activity, improving temporal alignment accuracy; 2. Multi-dimensional data enhancement significantly improves the robustness of the model under noisy environments and individual differences; 3. Multi-scale spatiotemporal attention and bidirectional LSTM collaboratively capture key regions and long-term temporal dependencies of the waveform; 4. A four-dimensional loss function comprehensively constrains waveform morphology, phase, key features, and spectrum, giving the reconstructed signal clinical diagnostic value. Experiments show that the method of this invention achieves a Pearson correlation coefficient of 0.954, an R-wave localization error as low as 2.47 milliseconds, and a heart rate variability analysis error of less than 7%. The method of this invention achieves a high level in terms of overall waveform correlation, key feature point localization, and preservation of physiological rhythm information. The model's accurate R-wave localization is particularly outstanding, which is crucial for applications such as arrhythmia detection. Meanwhile, the excellent long sequence reconstruction capability and HRV analysis accuracy demonstrate that the method of the present invention can generate clinically valuable electrocardiograms suitable for continuous monitoring and in-depth analysis from non-contact radar signals, providing a reliable technical path for non-intrusive health monitoring.

[0166] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Various modifications and variations can be made to the invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the scope of the claims of the invention.

Claims

1. A non-contact ECG generation method based on radar signals and enhancement strategies, the specific steps of which are as follows: S1. Perform data collection and preprocessing to construct a paired dataset; First, the synchronously acquired radar signals and corresponding electrocardiogram signals are used as training and testing data for the model. in, The radar signal is the I / Q signal acquired by a continuous wave radar system; the electrocardiogram signal is a synchronous contact ECG record with clinical reference value, i.e., a reference ECG signal. Then, the radar I / Q signals are preprocessed, that is, elliptic correction is achieved by fitting the I / Q trajectory to the ideal unit circle. The generalized quadratic curve model expression is as follows: ； in, and These represent the in-phase component and quadrature component of the radar I / Q signal, respectively. , , , , , All coefficients are real numbers, and , , Determines the shape of the trajectory. , Reflects the DC bias of the signal. For constant terms; Reconstruct the data matrix , its first Behavior ; Indicates the first The values ​​of the in-phase components at each data sampling point correspond to the I component in the raw I / Q signal received by the radar. Indicates the first The values ​​of the orthogonal components of each data sampling point correspond to the Q component in the original radar I / Q signal; Then solve the matrix equation. ,in The coefficient vector corresponding to the minimum singular value is obtained by using singular value decomposition (SVD). The least squares estimation of the model parameters is achieved; then, based on the solved coefficients, an inverse correction transformation is constructed to correct the distorted I / Q trajectory and map it to the vicinity of the ideal unit circle around the origin. in, This represents the matrix transpose operation; Then, the radar heart sound signal is extracted from the corrected phase signal, the heart sound is separated by 0–80 Hz bandpass filtering, and aligned with the preprocessed reference ECG signal on the time axis. Among them, the ECG signal is preprocessed, that is, the wavelet threshold denoising algorithm is used to suppress power frequency interference, electromyographic noise and baseline drift, improve the signal-to-noise ratio and retain the key waveform morphology. The key waveform morphologies include: P-wave, QRS complex, and T-wave; Finally, the preprocessed long-time-series signal, namely the preprocessed radar heart sound signal and the reference ECG signal, is segmented into segments of a preset fixed length to ensure that each segment includes 2–3 complete cardiac cycles. The preprocessed radar heart sound signal and the reference ECG signal are then time-aligned and amplitude-normalized to form "radar signal – ECG signal" paired samples. The final dataset contains N pairs of paired samples, which are divided into training set, validation set and test set according to a preset ratio. Among them, multi-dimensional data augmentation is performed on the training set samples before the training phase begins, including: temporal pruning and interpolation, time axis scaling, amplitude scaling, frequency domain selective filtering, and adaptive Gaussian noise injection. S2. Based on step S1, construct the RAE-ECG generation model; Construct a deep sequence generation model, namely the RAE-ECG generation model, which includes three cascaded modules: a feature extraction encoder, a temporal modeling module, and a sequence generation decoder; S3. Based on the RAE-ECG generation model constructed in step S2, input the paired dataset constructed in step S1, train and test the model, and output the reconstructed ECG. First, the training set is input into the RAE-ECG generation model for model training. That is, the pre-processed radar heart sound signal is used as input, the synchronously collected real electrocardiogram signal is used as the supervision target, and the model is optimized in an end-to-end manner. The training objective of the RAE-ECG generative model employs a four-dimensional composite loss function. The driver, this function, is driven by weighting coefficients. , , , Coordinate the four sub-items to simultaneously optimize the time-domain waveform similarity of the reconstructed signal. Phase consistency Accuracy of localization of key diagnostic features and frequency domain energy distribution The expression is as follows: ； The key diagnostic features include: P wave, R wave, and T wave; The RAE-ECG generative model is then trained using a mini-batch gradient descent strategy, with specific settings including: The optimizer used is AdamW, and the weight decay coefficient λ > 0 is set. The gradient accumulation strategy is adopted, and the weights are updated once every k steps. The learning rate scheduling adopts the OneCycle strategy, which linearly increases to the peak η_max > 0 in the first r% of training steps, and 0 < r < 100. The remaining steps are decayed by the cosine annealing function. The total number of training rounds is set to T_max, and a dynamic early stopping mechanism is introduced. That is, the validation set is input for validation. After training starts for w rounds, if the relative improvement of the validation set loss for p consecutive rounds does not exceed the threshold δ > 0, the training will be automatically terminated. After the RAE-ECG generation training is completed, the model performance is evaluated using an independent test set. The radar signal segments of the test set are input into the trained model, that is, sequentially input into the feature extraction encoder, the time series modeling module, and the sequence generation decoder, and the corresponding ECG waveform segments are automatically reconstructed. In order to evaluate the practicality of the model in long-term heart rhythm analysis, an overlapping segmentation and weighted merging strategy is adopted to reconstruct continuous ECG signals from the segments. That is, the original long sequence radar signal is segmented into segments with 50% overlap. After each segment is reconstructed independently, the overlapping area is weighted and averaged using a Hanning window to finally generate a continuous and complete high-fidelity ECG signal. S4. Based on the reconstructed ECG output from step S3, a systematic evaluation is performed according to the evaluation indicators to obtain a clinically usable ECG, thus completing the non-contact to clinical-grade ECG generation. The evaluation metrics include: waveform correlation, key wave localization, spectral consistency, and HRV analysis. First, the overall waveform similarity of the reconstructed ECG was assessed. Then, time-frequency analysis was used to ensure that the reconstructed signal was consistent with the spectral characteristics of the real ECG in the clinical core frequency band. Next, the model's localization accuracy of key diagnostic features was evaluated, and its time synchronization and morphological integrity were confirmed. Finally, heart rate variability (HRV) analysis was performed on the reconstructed continuous ECG signal. The key diagnostic features include: P wave, R wave, and T wave; After multi-dimensional verification by evaluation indicators, the model stably converts the input non-contact radar signal into a standard ECG signal with clear waveform characteristics, complete time and frequency information, and usable for heart rhythm analysis, realizing the generation of ECG signals from non-contact to clinical level.

2. The non-contact ECG generation method based on radar signals and enhancement strategies according to claim 1, characterized in that, In step S2, the RAE-ECG generation model is specifically as follows: A1. Feature Extraction Encoder: Employs a hierarchical, multi-stage convolutional architecture to progressively refine the original waveform into a feature representation rich in multi-scale contextual information through continuous downsampling and feature transformation. The feature extraction encoder receives a preprocessed radar heart sound signal sequence as input. The encoding process consists of four stages: The first stage uses a one-dimensional convolutional kernel of length 15 for initial feature mapping and maintains the temporal length through mirror padding, increasing the number of channels to 64; The second and third stages each consist of two cascaded residual blocks, gradually increasing the number of channels to 128 and using max pooling with strides of 4 and 5 to compress the temporal dimension, thereby expanding the receptive field and suppressing noise; The fourth stage stacks three residual blocks, where the convolutional kernel size decreases progressively with sizes of 7, 5, and 3, and the number of channels is fixed at 256. Multi-scale feature fusion is achieved through the continuous stacking of small convolutional kernels; Finally, the feature extraction encoder outputs a high-level feature sequence with compressed temporal dimension and expanded channel dimension. Each residual block includes batch normalization and exponential linear unit activation, and alleviates gradient vanishing through skip connections, supporting deeper network stacking; the structural expression of each residual block is as follows: ； in, and These represent the input and output tensors of the residual block, respectively. This represents the weight set { Parameterized main path nonlinear transformation; Indicating bypass The weight matrix of the convolution is used to weight the input. Perform the necessary linear projection to match the dimensions; the operator '+' indicates element-wise addition; A2, Timing Modeling Module; The temporal modeling module adopts an architecture based on self-attention mechanism and bidirectional long short-term memory network to model the long-term dependencies within the feature sequence and focuses on key temporal segments related to ECG features. The temporal modeling module receives the feature sequence encoded by the feature extraction encoder, including: two multi-head attention modules and a bidirectional long short-term memory network (Bi-LSTM) module; First, the feature sequence encoded by the feature extraction encoder is input into the multi-head attention module. That is, the multi-head attention mechanism captures the long-range dependencies of the sequence, splitting the features into 8 heads, and each head independently calculates attention, as shown in the following expression: ； in, This represents the dimension of the key vector, used as a scaling factor to stabilize the gradient; Indicates the attention head index; This represents the normalization function along the last dimension; , , Let these represent the generated query matrix, key matrix, and value matrix, respectively, and satisfy the following expression: ； in, Represents the input feature matrix; They represent the first The corresponding trainable query, key, and value projection weight matrix for each head; The outputs of all attention heads in each multi-head attention module are concatenated and linearly transformed to obtain the final output of the multi-head attention module. Then, the Bi-LSTM (Bi-Long Short-Term Memory) network module receives the sequence after attention weighting by all multi-head attention modules, and uses bi-directional recursive connections to accurately model the quasi-periodic patterns and long-range temporal dependencies within the heartbeat cycle, as shown in the following expression: ； ； ； in, Indicates the time step index of the sequence; Indicates time step The input feature vector; and These represent the forward and backward LSTM units at time steps, respectively. The hidden state; and These represent the hidden states at the previous time step in the corresponding direction; Represents a computation function that includes gating mechanisms; operators This represents a vector concatenation operation; the LSTM unit regulates the information flow through the forget gate, input gate, and output gate. A3, Sequence Generation Decoder; The sequence generation decoder receives the high-level feature representation after attention weighting and temporal modeling in the temporal modeling module; the decoder adopts a cascaded architecture of hierarchical upsampling and residual refinement; First, the feature tensor is processed by a local connection layer to smooth the feature distribution and suppress high-frequency noise. Then, it enters the first stage of upsampling, where the sequence length is extended to 1250 points based on the principle of spectrum shifting, and a double residual block is embedded to filter and optimize the extended signal. Then, it enters the second stage of upsampling, where the signal is stretched to a full length of 5000 points, and a physiological feature guidance mechanism is introduced. That is, by synergistically using multi-scale convolution kernels, the reconstruction process conforms to the physiological laws of ECG signals, and finally outputs a standardized ECG signal sequence that is clinically usable in both time domain morphology and frequency domain distribution.

3. The non-contact ECG generation method based on radar signals and enhancement strategies according to claim 1, characterized in that, In step S3, the four-dimensional composite loss function is as follows: Time-domain waveform similarity of reconstructed signals That is, amplitude loss, which is addressed by introducing the attention weight matrix of the ECG R wave. The weighted sum is expressed as follows: ； in, Indicates the amplitude loss value; This represents the expected value, and the batch average is taken in actual training. and These represent the predicted waveform of the non-contact ECG signal and the actual contact ECG reference waveform, respectively. This represents the absolute value operation; Phase consistency Phase loss, measured by first-order difference cosine similarity, is expressed as follows: ； in, This represents the phase loss value; Indicates the total length of the electrocardiogram (ECG) signal sequence; Indicates the time step index; The first-order difference operator is defined as follows: It is used to capture instantaneous changes in signals; and These represent the time steps. The predicted ECG signal value and the actual ECG signal value; This represents the function for calculating cosine similarity. R-wave positioning accuracy That is, the R-wave focusing loss, and the L1 loss calculated in the critical region, as shown in the following expression: ； in, Represents the R-wave focusing loss value; A represents the ECG R-wave attention weight matrix; Frequency domain energy distribution That is, the frequency domain loss, combined with the short-time Fourier transform amplitude and phase error, is expressed as follows: ； ； ； in, This represents the total frequency domain loss value; and These represent the amplitude spectrum loss and the phase spectrum loss, respectively. Represents the short-time Fourier transform; Indicated in frequency index and time frame index Amplitude spectrum value at; and Indicates the corresponding phase value; and These represent the total number of frequency domain cells and the total number of time domain frames, respectively. and Indicates the summation index; This represents summing over all time-frequency units; This represents the normalization factor for calculating the average. 0.5 represents the sine function; 0.5 is the weighting coefficient for phase loss.

Images