Millimeter wave electrocardiogram reconstruction system and method based on FDE-MSTA-DMF cross-domain diffusion model

Abstract

The invention discloses a millimeter wave electrocardiogram reconstruction system and method of a cross-domain diffusion model based on FDE-MSTA-DMF, and belongs to the technical field of non-contact electrocardiogram monitoring. The system comprises a millimeter wave radar acquisition module, a mechanical domain signal preprocessing module, a cross-domain diffusion reconstruction module, a frequency domain feature enhancement module, a multi-scale time sequence attention module, a dynamic multi-modal fusion module and an electrocardiogram output module. 0.5-40Hz layered frequency domain feature extraction of ECG signals, multi-scale time sequence modeling of short-time QRS wave dynamic and long-time heartbeat cycle rhythm and millimeter wave-ECG dynamic weight fusion based on cosine similarity are realized. The method aims at normal heartbeat scenes of healthy people, solves the problems of lack of frequency domain features, single time sequence modeling and stiffness of multi-modal fusion in the prior art, and is suitable for non-inductive continuous electrocardiogram monitoring application such as intelligent medical treatment and home health monitoring.

Description

Technical Field

[0001] This invention relates to the fields of biomedical engineering, radar signal processing and artificial intelligence, and in particular to a millimeter-wave electrocardiogram reconstruction system and method based on the FDE-MSTA-DMF cross-domain diffusion model, which is suitable for non-contact ECG monitoring of normal heartbeats in healthy individuals. Background Technology

[0002] Cardiovascular disease remains one of the leading causes of death worldwide, and continuous, accurate electrocardiogram (ECG) monitoring is crucial for early diagnosis and intervention. While traditional contact ECGs offer high signal quality, they suffer from skin irritation, electrode displacement, and patient discomfort, limiting their application in long-term monitoring, neonatal care, or burn wards. Millimeter-wave (mmWave) radar, as a non-contact sensing method, offers advantages such as penetrating clothing, resistance to light interference, and privacy protection, providing an ideal solution for unobtrusive ECG monitoring.

[0003] Existing technologies still have three core limitations: First, insufficient utilization of frequency domain features. ECG signals contain layered characteristics of low-frequency heart rate (0.5-4Hz) and mid-to-high-frequency morphology (4-40Hz), but existing models only process in the time domain, resulting in a lack of physiological realism in the generated signals. Second, limited capture of multi-scale temporal dependencies. A normal heartbeat contains short dynamics of QRS waves (200-300ms) and long-period rhythms (0.8-1.2s). Single-scale attention mechanisms cannot simultaneously adapt to both types of dynamics, resulting in large RR interval errors. Third, low efficiency of multimodal fusion. The correlation between millimeter-wave signals and ECG signals changes dynamically with factors such as human posture and radar distance. Fixed-weight fusion cannot adaptively adjust, which may introduce redundant noise or lose key information.

[0004] To address the aforementioned issues, this invention focuses on non-contact ECG reconstruction of normal heartbeats in healthy individuals and proposes an improved cross-domain diffusion Transformer architecture. Through three innovative modules—frequency domain feature enhancement, multi-scale temporal attention, and dynamic multimodal fusion—it significantly improves reconstruction accuracy, morphological fidelity, and temporal accuracy. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a millimeter-wave electrocardiogram reconstruction system based on the FDE-MSTA-DMF cross-domain diffusion model. This method is based on millimeter-wave radar for non-contact electrocardiogram reconstruction and is particularly suitable for normal heartbeat scenarios in healthy individuals.

[0006] The millimeter-wave electrocardiogram reconstruction system based on the cross-domain diffusion model of FDE-MSTA-DMF provided by this invention includes a millimeter-wave radar acquisition module, a mechanical domain signal preprocessing module, a cross-domain diffusion reconstruction module, a frequency domain feature enhancement module, a multi-scale temporal attention module, a dynamic multimodal fusion module, and an electrocardiogram output module. The millimeter-wave radar acquisition module is used to acquire the raw echo signal of the target's chest cavity region using millimeter-wave radar; The mechanical domain signal preprocessing module is used to perform phase calculation and acceleration filtering on the echo signal to obtain an effective physiological vibration signal sequence. The cross-domain diffusion reconstruction module achieves electrocardiogram reconstruction through a forward noise diffusion and reverse denoising generation mechanism; The frequency domain feature enhancement module is used to perform fast Fourier transform on the pure electrocardiogram signal, extract low-frequency heart rate features and high-frequency morphological features in layers, and embed the frequency domain information into the diffusion process. The multi-scale temporal attention module is used to capture the short-timescale QRS wave dynamics and long-timescale heart rhythm of electrocardiogram signals through a dual-scale convolution and attention fusion mechanism. The dynamic multimodal fusion module is used to dynamically allocate fusion weights between millimeter-wave signals and electrocardiogram signals based on feature similarity. The electrocardiogram output module is used to perform regression mapping on the feature vector output by the cross-domain diffusion reconstruction module and output the normalized reconstructed electrocardiogram signal.

[0007] Furthermore, the mechanical domain signal preprocessing module includes a phase calculation unit, an acceleration filtering unit, and a multi-channel patching unit; The phase calculation unit is used to calculate the chest wall vibration displacement based on the phase of the reflected signal. The calculation formula is as follows: ,in λ For radar wavelength, for Phase of the reflected signal at any moment; The acceleration filtering unit is used to eliminate low-frequency breathing noise through second-order difference while retaining the high-frequency vibration components caused by cardiac contraction / diastole. The filtering formula is as follows: Where h is the sampling step size; The multi-channel patching unit is used to reshape a one-dimensional time-series signal into a two-dimensional feature map and add position encoding to generate mechanical domain conditional embedding features.

[0008] Furthermore, the cross-domain diffusion reconstruction module includes a forward diffusion unit, a noise prediction unit, and a reverse denoising unit; The forward diffusion unit is used to gradually add Gaussian noise to the pure electrocardiogram signal to construct a Markov degradation process from pure signal to pure noise, with a total diffusion step count T=1000. The noise prediction unit is used to predict the noise distribution at each diffusion step using a hierarchical Transformer architecture. The Transformer contains 6 blocks, each of which consists of adaptive layer normalization, multi-head self-attention and a feedforward network. The reverse denoising unit is used to gradually remove noise based on the predicted noise distribution and reconstruct the target electrocardiogram signal.

[0009] Furthermore, the frequency domain feature enhancement module includes an FFT transform unit, a frequency band separation coding unit, and a frequency domain fusion unit; The FFT transform unit is used to perform a fast Fourier transform on the pure electrocardiogram signal to obtain frequency domain amplitude and phase information; The frequency band separation coding unit is used to divide the frequency domain signal into a 0.5-4Hz heart rate band and a 4-40Hz morphological band by resampling 64 frequency bins. The heart rate branch uses a 1D convolution with kernel size=21 and the morphological branch uses a 1D convolution with kernel size=7 to extract layered features respectively. The frequency domain fusion unit is used to combine frequency domain features with dynamic weights. γ =0.05 is superimposed on the noise prediction process to constrain the elimination of high-frequency morphological noise. The corrected noise estimate is: .

[0010] Furthermore, the multi-scale temporal attention module includes a dual-scale convolution unit, an attention computation unit, and a feature fusion unit; The dual-scale convolutional unit is used to capture short dynamics of QRS waves with convolution of kernel size=3 and to capture long-cycle rhythm of heartbeats with convolution of kernel size=7. Both use grouped convolution to reduce computational complexity. The attention calculation unit is used to calculate the attention weight of short-scale features on long-scale features, highlighting the contribution of key time nodes of R-wave and T-wave. The formula for calculating the attention weight is as follows: Where D=384 is the feature dimension. This is a scaling factor to avoid excessive attention weights; The feature fusion unit is used to superimpose multi-scale temporal features onto the Transformer temporal features with a dynamic weight β=0.1, resulting in the following updated features: .

[0011] Furthermore, the dynamic multimodal fusion module includes a similarity calculation unit, a dynamic weight allocation unit, and a conditional feature update unit; The similarity calculation unit is used to calculate the global cosine similarity between millimeter-wave conditional embedding features and electrocardiogram time-domain features, and to measure the correlation strength between the two modalities. The dynamic weight allocation unit is used to map similarity to the dynamic fusion weights in the range [0.3, 0.7] using the sigmoid function. Add an offset δ=0.5 to avoid extreme weights; The conditional feature update unit is used to re-weight and fuse multimodal features according to dynamic weights to update the mechanical domain conditional embedding: .

[0012] Furthermore, it also includes a model training module for training the model based on the Schellenberger Clinical Radar Vital Signs Dataset, which contains 86,459 seconds of synchronized data from 30 healthy subjects, with 91.3% of the samples exhibiting normal sinus rhythm at rest. A multi-task loss function is constructed using L1 loss, mean squared error, frequency domain loss, temporal loss, and fusion loss. Weight parameters =0.05、 =0.1、 =0.02 was determined through grid search; an early stopping strategy was introduced and the AdamW optimizer was used for iterative training with an initial learning rate of 5e-5, a batch size of 16, and a total of 70 epochs.

[0013] Furthermore, it also includes an output evaluation module, which is used to comprehensively evaluate the reconstructed electrocardiogram using Pearson correlation coefficient, mean square error, frequency domain correlation coefficient, R wave detection accuracy, and RR interval standard deviation; among which, the R wave detection accuracy reaches 0.9638, the RR interval standard deviation is reduced to 3.9951ms, and the frequency domain correlation coefficient reaches 0.9968.

[0014] The millimeter-wave electrocardiogram reconstruction system based on the FDE-MSTA-DMF transdomain diffusion model provided by this invention includes the following steps: S1: Use millimeter-wave radar to collect echo signals from the target's thoracic cavity region, and extract pure mechanical domain condition signals through phase calculation and acceleration filtering; S2: Perform multi-channel patching processing on the mechanical domain condition signal to generate embedded features that match the dimensions of the electrocardiogram signal; S3: Perform a positive noise diffusion process on the pure electrocardiogram signal to generate a noisy electrocardiogram sequence, and fuse it with the mechanical domain condition embedding to form diffusion conditions; S4: In the reverse denoising process, noise prediction is optimized sequentially through the frequency domain feature enhancement module, the multi-scale temporal attention module, and the dynamic multimodal fusion module; S5: Substitute the final predicted noise into the denoising formula to gradually reconstruct the pure electrocardiogram signal and output it.

[0015] Furthermore, the frequency domain feature enhancement module operates in the reverse denoising process, specifically including: S41: Perform an FFT transform on the current step's pure electrocardiogram estimate and calculate the power spectral density: ; S42: The frequency domain signal is divided into a low-frequency band for heart rate and a mid-to-high-frequency band for morphology. Features are extracted from each band using dedicated convolutional branches, and then compressed into 384-dimensional frequency domain fusion features via linear mapping. ; S43: Weight the frequency domain fusion features =0.05 is added to the noise prediction output to correct the noise estimate and update the denoised mean.

[0016] Furthermore, the multi-scale temporal attention module operates on the inverse denoising process, specifically including: S44: Temporal characteristics of the Transformer output Parallel grouped convolutions with kernel size=3 and kernel size=7 are performed to extract short-scale features. With long-scale features ; S45: Calculate the attention weight matrix of short-scale features on long-scale features, and fuse them to generate multi-scale temporal features. ; S46: Weight multi-scale temporal features =0.1 is superimposed on the original time-domain features to update the noise predictor input. .

[0017] Furthermore, the dynamic multimodal fusion module operates in the reverse denoising process, specifically including: S47: Real-time computation of millimeter-wave conditional embedding With electrocardiogram time domain characteristics Global cosine similarity; S48: Map similarity to dynamic fusion weights using the sigmoid function. When the similarity of the resting scene is approximately 0.87 ≈0.68, when the similarity of the breathing interference scene is ≈0.52. ≈0.58; S49: Re-weight and fuse multimodal features according to the weights, and update the mechanical domain conditional embedding. .

[0018] The beneficial effects of this invention are as follows: This invention provides a millimeter-wave electrocardiogram (ECG) reconstruction system based on a cross-domain diffusion model using FDE-MSTA-DMF, overcoming the limitations of traditional contact-based ECG acquisition and the shortcomings of existing non-contact reconstruction methods in terms of morphological detail, temporal accuracy, and robustness. The method first extracts pure mechanical domain conditional signals through millimeter-wave radar signal acquisition and preprocessing; secondly, it utilizes a cross-domain diffusion framework to construct a mapping relationship from millimeter-wave to ECG. The core innovations lie in three aspects: the frequency domain feature enhancement module supplements layered frequency information through FFT and frequency band coding, solving the problem of morphological detail loss and reducing T-wave amplitude deviation by 25.7%, achieving a frequency domain correlation coefficient of 0.9968; the multi-scale temporal attention module simultaneously captures short-dynamic QRS waves and long-cycle heartbeats through dual-scale convolution and attention fusion, reducing the RR interval error to 3.9951ms, a 61.2% reduction compared to the original model; and the dynamic multimodal fusion module adaptively adjusts fusion weights based on cosine similarity, increasing the qualified monitoring rate to 92% under respiratory interference scenarios, a 5.35% improvement compared to a fixed fusion scheme.

[0019] This method features non-contact and comfort: completely avoiding traditional electrode patches, suitable for long-term continuous monitoring; joint frequency-time domain modeling: fully utilizing the frequency layering characteristics of ECG to improve morphological fidelity; multi-scale time-series adaptation: accurately capturing the dynamics of the heartbeat across multiple time scales, optimizing time-series accuracy; dynamic fusion robustness: adaptively responding to changes in modal correlation, enhancing anti-interference capabilities. Experimental validation shows that on the Schellenberger clinical dataset, this method achieves a Pearson correlation coefficient of 0.9863, a mean squared error of 0.0274, an R-wave detection accuracy of 0.9638, a 42.9% reduction in model MSE, and a 61.2% reduction in RR interval time-series error. It outperforms traditional baseline models in waveform reconstruction accuracy, signal time-series consistency, and model generalization ability, making it suitable for various applications such as smart healthcare, home health monitoring, and remote monitoring.

[0020] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0021] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the following drawings are provided for illustration.

[0022] Figure 1 This is the overall system architecture diagram.

[0023] Figure 2 This is a flowchart of the preprocessing process for signals in the mechanical domain.

[0024] Figure 3 This is a structural diagram of the frequency domain feature enhancement module.

[0025] Figure 4 This is a structural diagram of a multi-scale temporal attention module.

[0026] Figure 5 This is a structural diagram of the dynamic multimodal fusion module. Detailed Implementation

[0027] Specific embodiments are provided to further illustrate the present invention so that those skilled in the art can better understand and implement it, but the embodiments are not intended to limit the present invention.

[0028] Example 1 like Figure 1 As shown, Figure 1 The system overall structure diagram provided in this embodiment is a millimeter-wave electrocardiogram reconstruction system based on the cross-domain diffusion model of FDE-MSTA-DMF. It includes a millimeter-wave radar acquisition module, a mechanical domain signal preprocessing module, a cross-domain diffusion reconstruction module, a frequency domain feature enhancement module, a multi-scale temporal attention module, a dynamic multimodal fusion module, an electrocardiogram output module, a model training module, and an output evaluation module.

[0029] The millimeter-wave radar acquisition module uses a 24GHz six-port continuous wave radar with a dual-base antenna design, a focal length of 40cm, a phase resolution of 0.1°, a sampling frequency of 2000Hz, and an ADC resolution of 24 bits. It is deployed within a range of 30-60cm from the subject's chest.

[0030] like Figure 2 As shown, Figure 2 As shown in the system overall structure diagram, the mechanical domain signal preprocessing module processes the raw signals captured by the millimeter-wave radar through the following signal processing flow: First, the chest wall vibration region is located by searching the round-trip length (RTL) profile (selecting the region with the highest vibration energy), and the phase of the reflected signal is extracted. Subsequently, the chest wall vibration displacement was calculated. Then, breathing noise is eliminated by acceleration filtering to obtain a pure mechanical domain condition signal. Finally, multi-channel patching was used to... From 1D timing signals ( (8 represents the number of virtual channels) reshaped into a 2D feature map ( The mechanical domain conditional embedding is obtained by using a CNN encoder (kernel size=2, stride=2, output dimension 384) and 2D position encoding. This provides mechanical domain supervision for the subsequent diffusion process.

[0031] The cross-domain diffusion reconstruction module has a total diffusion step count T=1000 and a scale parameter. Forward diffusion occurs through cumulative product. Directly generate noisy ECG: , The noise predictor is a 6-layer Transformer, with each layer containing 8 self-attention mechanisms, a hidden dimension of 384, and a feedforward network expansion factor of 4.

[0032] like Figure 3 As shown, Figure 3 The diagram shows the structure of the frequency domain feature enhancement module. The physiological information of the ECG signal from this module exhibits clear frequency domain stratification: the low-frequency band (0.5-4Hz) corresponds to heart rate and heart rate variability (HRV), reflecting cardiac rhythm stability; the mid-to-high frequency band (4-40Hz) corresponds to morphological features such as the QRS complex (10-20Hz) and T wave (0.1-1Hz, with actual effective energy concentrated at 4-8Hz), directly determining the clinical diagnostic value of ECG. The original model only modeled using time-domain features, resulting in the loss of mid-to-high frequency morphological information. FDEM supplements this stratified frequency information through joint frequency-time domain processing.

[0033] The core process of the frequency domain feature enhancement module is "frequency domain feature extraction - frequency band separation coding - time domain fusion optimization": First, the pure ECG is processed... (flattened into a 1D sequence) Perform a Fast Fourier Transform (FFT) to obtain the frequency domain signal. Its amplitude and phase are as follows: , To focus on the key ECG frequency band (0.5-40Hz), Resampling was performed (from 512 frequency bins to 64, denoted as...) ), and calculate the power spectral density (PSD) to enhance the physiological significance of the frequency characteristics: Subsequently, two dedicated convolutional branches were designed to extract low-frequency heart rate features and mid-to-high-frequency morphological features, respectively: the heart rate branch uses a 1D convolution with kernel size=21 (covering the time window corresponding to the 0.5-4Hz frequency band, approximately 200ms), and obtains heart rate features through adaptive average pooling (output dimension 32) and linear mapping (input 32, output 384). The morphological branch uses a 1D convolution with kernel size=7 (covering the time window corresponding to the 4-40Hz frequency band, approximately 60ms), followed by pooling (output dimension 64) and linear mapping to obtain morphological features. To incorporate frequency domain features into a noise predictor for a cross-domain diffusion model, , and (Linearly mapped to 384 dimensions) spliced ​​together, and compressed into 384-dimensional frequency domain fusion features through an MLP with LayerNorm and GELU activation. and with dynamic weights (After grid search optimization, the contributions of frequency domain and time domain features are balanced) and then superimposed onto the input features of the original noise predictor to correct the noise estimation: Substituting the corrected noise estimate into the denoising mean formula, we obtain the denoising mean of the fused frequency domain information: This component enables the noise predictor to accurately identify mid-to-high frequency morphological noise through frequency domain feature constraints.

[0034] like Figure 4 As shown, Figure 4 This is a structural diagram of the multi-scale temporal attention module. The multi-scale temporal attention module addresses the significant multi-scale characteristics of the temporal dynamics of a normal heartbeat: short timescales (200-300ms) correspond to the ventricular depolarization process of the QRS wave, with dramatic signal changes; long timescales (0.8-1.2s) correspond to the periodic rhythm of continuous heartbeats, reflecting the stability of cardiac activity. The original model's single-scale Transformer attention cannot simultaneously adapt the global receptive field to both types of dynamics, resulting in a large RR interval error. MSTAM achieves accurate modeling of multi-scale temporal correlations by fusing dual-scale convolution with attention. The input to MSTAM is the temporal features output from the original layer's Transformer. (B represents the batch size, 256 represents the number of patches, and 384 represents the feature dimension). The core process is "multi-scale feature extraction - attention fusion - temporal optimization": First, two parallel 1D grouped convolutions (number of groups = 96, reducing computational complexity) are used to extract dual-scale features. The short-scale convolution (kernel size = 3, stride = 1) focuses on the local dynamics of the QRS wave, and the long-scale convolution (kernel size = 7, stride = 1) captures the global rhythm of the heartbeat. The formula is expressed as: , In the formula, It is a short-scale feature. It is a long-scale feature. , These are the kernel weights for the two types of convolutions, respectively. These correspond to batch, patch, and feature dimension index, respectively. To strengthen the correlation of multi-scale features, the attention weight of short-scale features on long-scale features is calculated, highlighting the contributions of key time-series nodes such as R-wave and T-wave: In the formula, For feature dimension, This is a scaling factor to avoid excessive attention weights. The fused temporal features are: LayerNorm ensures stable feature distribution and avoids gradient explosion. It fuses temporal features. With dynamic weights (Experimentally verified, balancing multi-scale and original temporal features) is superimposed onto the original Transformer features to update the input features of the noise predictor: Noise predictor based on Output corrected noise estimation Substituting into the denoising mean formula, we get: This component enables the model to simultaneously capture the short dynamics of the QRS wave and the long cycle of the heartbeat through multi-scale temporal modeling.

[0035] like Figure 5 As shown, Figure 5 The diagram shows the structure of the multi-scale temporal attention module. The dynamic multimodal fusion module dynamically allocates fusion weights based on cosine similarity, adapting to real-time changes in modal relevance and ensuring efficient integration of multimodal information. The core process of the DMFM component is "modal relevance calculation - dynamic weight allocation - conditional feature update": first, the millimeter-wave conditional embedding is calculated. ECG time domain characteristics The global cosine similarity measures the correlation strength between two modalities: In the formula, These correspond to batch and patch indices, respectively, with similarity values ​​ranging from -1 to 1. Higher values ​​indicate stronger modal associations. The sigmoid function maps the similarity to fusion weights (by adding an offset). To ensure that the weights are reasonably distributed within the range of 0.3-0.7, and to avoid extreme weights. in, The fusion weights for millimeter-wave features, The fusion weights for ECG features are used when modal similarity is high (e.g., in a resting scene, sim≈0.87). This enhances the mechanical-electrical mapping guidance of millimeter waves; when modal similarity is low (e.g., in breathing interference scenarios, sim≈0.52), To suppress noise, the weights of ECG features are increased. Dynamic weights are then applied to modal feature fusion to update the mechanical domain conditional embedding. Will Substituting the original noise predictor, we obtain the corrected noise estimate. The corresponding denoising mean is: This component adapts to modal correlation fluctuations through dynamic fusion.

[0036] The ECG output module regresses and maps the final denoised features through a fully connected layer (256→1024→1024→1024), constrains the output to the [-1,1] interval (normalized ECG voltage range) by the Tanh activation function, and outputs 1024 sampling points for a 5s segment.

[0037] The model training module is based on the Schellenberger dataset (30 healthy subjects, 86459 seconds, 91.3% resting NSR samples), with a data partition of 8:1:1. Training employs a multi-task loss mechanism. Weight , , All were determined through grid search to ensure a balanced contribution from each loss term. For the power spectrum MSE, For the RR interval MSE, =1−sim 。 AdamW optimizer (lr=5e-5, weight_decay=1e-4), cosine annealing warm restart, batch size 16 (gradient accumulation 2 steps), 70 epochs, early stopping patience=5.

[0038] The output evaluation module uses seven core metrics: Pearson_Corr=0.9863, Freq_Corr=0.9968, MSE=0.0274, RMSE=0.1642, RR_Interval_Std_Diff=3.9951ms, and R_Wave_Accuracy=0.9638. R-wave detection uses the Neurokit2 algorithm, and the QRS width is used to calculate the full width at half maximum (FWHM) of the waveform.

[0039] Example 2 This embodiment is a specific implementation of the millimeter-wave electrocardiogram reconstruction system method based on the FDE-MSTA-DMF cross-domain diffusion model: S1: Acquire echo signals from the target chest cavity region using a 24GHz millimeter-wave radar at a frame rate of 250fps. Locate the chest wall vibration region using RTL profiling and extract the phase. Calculate displacement Accelerometer filtering eliminates respiratory noise while preserving high-frequency components related to heartbeat, thus obtaining the mechanical domain condition signal. .

[0040] S2: Yes Multi-channel patching was performed, reshaping the B×8×1024 to B×8×32×32. The CNN encoder extracted 384-dimensional features and added 2D positional encoding to generate... The synchronously acquired ECG true value X0 (sampled at 1000Hz) was resampled to 1024 points and patched in 32×32 format. .

[0041] S3: Forward diffusion random sampling t∈[1,1000] generates noisy ECG. ,and Fusion forms diffusion conditions .

[0042] S4: Reverse denoising iterates from t=1000 to t=1: Noise predictor output initial and ; The FDE module performs an FFT on the current X0 estimate to extract 64-bin frequency domain features and corrects them. ; MSTAM module Perform dual-scale convolution and attention fusion, and update ; The DMFM module calculates similarity and updates. ; Final prediction noise Calculate the denoised mean ,sampling .

[0043] S5: When hour, Output pure ECG reconstruction results After Tanh normalization, it is brought to the standard range of [-1,1]mV.

[0044] Example 3 The dataset used was the Schellenberger Clinical Radar Vital Signs Dataset, containing synchronous millimeter-wave and ECG data from 30 healthy subjects (age 30.7±9.9 years, BMI 23.2±3.3 kg / m²). The millimeter-wave device was a 24 GHz six-port continuous-wave radar (bistatic antenna, focal length 40 cm, phase resolution 0.1°). ECGs were recorded synchronously via three leads (sampling rate 1000 Hz, accuracy ±5 μV). 91.3% of the samples showed normal sinus rhythm (NSR) at rest, with a total recording time of 86459 s (approximately 24 h). Data preprocessing workflow: Millimeter wave signals are normalized at the channel level and bandpass filtered from 0.5 to 40 Hz. ECGs are segmented into 5-second segments (1024 sampling points), and the training set / validation set / test set are divided in an 8:1:1 ratio. During the training phase, the sample size is increased by 2 times by using a time offset of ±20 sampling points and Gaussian noise with a standard deviation of 0.02.

[0045] The evaluation metrics cover morphology, temporal sequence, frequency domain, and clinical efficacy: morphological similarity is measured using Pearson correlation coefficient (Pearson_Corr), Spearman correlation coefficient (Spearman_Corr), cross correlation coefficient (CrossCorr), mean squared error (MSE), root mean square error (RMSE), and mean absolute error (MAE); temporal accuracy is measured using QRS width difference (QRS_Width_Diff) and RR interval standard deviation difference (RR_Interval_Std_Diff); frequency domain consistency is measured using frequency domain correlation coefficient (Freq_Corr); and clinical efficacy is measured using R-wave detection accuracy (R_Wave_Accuracy). All metrics are calculated using Neurokit2 and custom scripts.

[0046] Experimental environment: Hardware consisted of an NVIDIA GTX 1660ti GPU (6GB VRAM), an Intel i7-10900K CPU, and 64GB RAM; Software consisted of Python 3.8, PyTorch 1.10.0, and CUDA 11.3; Training parameters: Batch size 16 (gradient accumulation 2 steps equivalent to batch size 32), AdamW optimizer (weight decay 1e-4, betas=(0.9, 0.999)), initial learning rate 5e-5 (cosine annealing warm restart, 5 epochs preheating), total epochs 70, diffusion steps 1000, early stopping policy (patience=5, based on the validation set Pearson_Corr), automatic mixed precision training (use_amp=True) and gradient clipping (threshold 0.5) enabled.

[0047] This embodiment is for ablation experiment verification, with 5 control groups configured: Baseline: PCC=0.9550, MSE=0.0480, RR_Interval_Std_Diff=10.3000ms; Removing FDE: PCC decreased to 0.9721, Freq_Corr decreased to 0.9853; Remove MSTAM: RR_Interval_Std_Diff increased to 5.5280ms, R_Wave_Accuracy decreased to 0.9412; Removed DMFM: QRS_Width_Diff increased to 0.1967, R_Wave_Accuracy decreased to 0.9386; The full component model (in this invention) has the following optimal values: PCC=0.9863, MSE=0.0274, RR_Interval_Std_Diff=3.9951ms, and R_Wave_Accuracy=0.9638.

[0048] Example 4 This embodiment is a comparative experiment for verification, comparing with four types of baseline models: RF-ECG (cGAN): PCC=0.8640, R_Wave_Accuracy=0.6840; CTL-ECG (CNN): PCC=0.8900, R_Wave_Accuracy=0.7280; DR-MUSIC (BiLSTM): PCC=0.9420, R_Wave_Accuracy=0.8350; AirECG (diffusion): PCC=0.9550, R_Wave_Accuracy=0.9200; The improved model of this invention has leading performance with PCC=0.9863 and R_Wave_Accuracy=0.9638.

[0049] Example 5 This example illustrates the deployment and application. After training, the model is exported in ONNX format and deployed on an NVIDIA Jetson Nano edge device (GPU 128 CUDA cores, 4GB RAM). During inference, the latency per batch (5s segment) is 82ms, and power consumption is less than 6W. It supports real-time monitoring of scenarios such as smart mattresses, office chairs, and nursing chairs, with inference latency below 100ms, meeting continuous monitoring requirements. Results are wirelessly transmitted to a cloud analysis platform via Bluetooth / 5G, adapting to home health monitoring and telemedicine applications.

[0050] The above-described embodiments are merely preferred embodiments provided to fully illustrate the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention. The scope of protection of the present invention is defined by the claims.

Claims

1. A millimeter-wave electrocardiogram reconstruction system based on the FDE-MSTA-DMF transdomain diffusion model, characterized in that, It includes a millimeter-wave radar acquisition module, a mechanical domain signal preprocessing module, a cross-domain diffusion reconstruction module, a frequency domain feature enhancement module, a multi-scale temporal attention module, a dynamic multimodal fusion module, and an electrocardiogram output module; The millimeter-wave radar acquisition module is used to acquire the raw echo signal of the target's chest cavity region using millimeter-wave radar; The mechanical domain signal preprocessing module is used to perform phase calculation and acceleration filtering on the echo signal to obtain an effective physiological vibration signal sequence. The cross-domain diffusion reconstruction module is used to construct an improved diffusion model and realize electrocardiogram reconstruction through a forward noise diffusion and reverse denoising generation mechanism. The frequency domain feature enhancement module is used to perform fast Fourier transform on the pure electrocardiogram signal, extract low-frequency heart rate features and high-frequency morphological features in layers, and embed the frequency domain information into the diffusion process. The multi-scale temporal attention module is used to capture the short-timescale QRS wave dynamics and long-timescale heart rhythm of electrocardiogram signals through a dual-scale convolution and attention fusion mechanism. The dynamic multimodal fusion module is used to dynamically allocate fusion weights between millimeter-wave signals and electrocardiogram signals based on feature similarity. The electrocardiogram output module is used to perform regression mapping on the feature vector output by the cross-domain diffusion reconstruction module and output the normalized reconstructed electrocardiogram signal.

2. The millimeter-wave electrocardiogram reconstruction system based on the FDE-MSTA-DMF transdomain diffusion model according to claim 1, characterized in that, The mechanical domain signal preprocessing module includes a phase calculation unit, an acceleration filtering unit, and a multi-channel patching unit. The phase calculation unit is used to calculate the chest wall vibration displacement through the phase of the reflected signal; The acceleration filtering unit is used to eliminate low-frequency breathing noise through second-order difference while retaining the high-frequency vibration components caused by cardiac contraction / diastole. The multi-channel patching unit is used to reshape a one-dimensional time-series signal into a two-dimensional feature map and add position encoding to generate mechanical domain conditional embedding features.

3. The millimeter-wave electrocardiogram reconstruction system based on the FDE-MSTA-DMF transdomain diffusion model according to claim 1, characterized in that, The cross-domain diffusion reconstruction module includes a forward diffusion unit, a noise prediction unit, and a reverse denoising unit; The forward diffusion unit is used to gradually add Gaussian noise to the pure electrocardiogram signal to construct a Markov degradation process from pure signal to pure noise. The noise prediction unit is used to predict the noise distribution at each diffusion step using a hierarchical Transformer architecture. The reverse denoising unit is used to gradually remove noise based on the predicted noise distribution and reconstruct the target electrocardiogram signal.

4. The millimeter-wave electrocardiogram reconstruction system based on the FDE-MSTA-DMF transdomain diffusion model according to claim 1, characterized in that, The frequency domain feature enhancement module includes an FFT transform unit, a frequency band separation coding unit, and a frequency domain fusion unit; The FFT transform unit is used to perform a fast Fourier transform on the pure electrocardiogram signal to obtain frequency domain amplitude and phase information; The frequency band separation coding unit is used to divide the frequency domain signal into a 0.5-4Hz heart rate band and a 4-40Hz morphological band by using the resampling frequency bin, and to extract layered features using different convolution kernels; The frequency domain fusion unit is used to superimpose frequency domain features with dynamic weights onto the noise prediction process to constrain the elimination of high-frequency morphological noise.

5. The millimeter-wave electrocardiogram reconstruction system based on the FDE-MSTA-DMF transdomain diffusion model according to claim 1, characterized in that, The multi-scale temporal attention module includes a dual-scale convolutional unit, an attention computation unit, and a feature fusion unit; The dual-scale convolutional unit is used to capture short dynamics of QRS waves using convolution with kernel size=3 and to capture long-cycle rhythms of the heartbeat using convolution with kernel size=7. The attention calculation unit is used to calculate the attention weight of short-scale features on long-scale features, highlighting the contribution of key time nodes of R-wave and T-wave. The feature fusion unit is used to superimpose multi-scale temporal features with dynamic weights onto Transformer temporal features to optimize the modeling of heartbeat temporal correlation.

6. The millimeter-wave electrocardiogram reconstruction system based on the FDE-MSTA-DMF transdomain diffusion model according to claim 1, characterized in that, The dynamic multimodal fusion module includes a similarity calculation unit, a dynamic weight allocation unit, and a conditional feature update unit; The similarity calculation unit is used to calculate the global cosine similarity between millimeter-wave conditional embedding features and electrocardiogram time-domain features, and to evaluate the correlation strength between the two modalities. The dynamic weight allocation unit is used to map similarity to fusion weights through the sigmoid function, increasing millimeter wave weights when similarity is high and increasing electrocardiogram weights when similarity is low. The conditional feature update unit is used to re-weight and fuse multimodal features according to dynamic weights to update the mechanical domain conditional embedding.

7. The millimeter-wave electrocardiogram reconstruction system based on the FDE-MSTA-DMF transdomain diffusion model according to claim 1, characterized in that, It also includes a model training module for training models based on the Schellenberger clinical radar vital signs dataset. It uses L1 loss, mean square error, frequency domain loss, temporal loss and fusion loss to jointly construct a multi-task loss function, introduces an early stopping strategy and uses the AdamW optimizer for iterative training.

8. The millimeter-wave electrocardiogram reconstruction system based on the FDE-MSTA-DMF transdomain diffusion model according to claim 1, characterized in that, It also includes an output evaluation module, which is used to comprehensively evaluate the reconstructed electrocardiogram using Pearson correlation coefficient, mean square error, frequency domain correlation coefficient, R wave detection accuracy, and RR interval standard deviation.

9. A millimeter-wave electrocardiogram reconstruction method based on the FDE-MSTA-DMF transdomain diffusion model, characterized in that, Includes the following steps: S1: Use millimeter-wave radar to collect echo signals from the target's thoracic cavity region, and extract pure mechanical domain condition signals through phase calculation and acceleration filtering; S2: Perform multi-channel patching processing on the mechanical domain condition signal to generate embedded features that match the dimensions of the electrocardiogram signal; S3: Perform a positive noise diffusion process on the pure electrocardiogram signal to generate a noisy electrocardiogram sequence, and fuse it with the mechanical domain condition embedding to form diffusion conditions; S4: In the reverse denoising process, noise prediction is optimized sequentially through the frequency domain feature enhancement module, the multi-scale temporal attention module, and the dynamic multimodal fusion module; S5: Substitute the final predicted noise into the denoising formula to gradually reconstruct the pure electrocardiogram signal and output it.

10. The millimeter-wave electrocardiogram reconstruction method based on the FDE-MSTA-DMF transdomain diffusion model according to claim 9, characterized in that, The frequency domain feature enhancement module, multi-scale temporal attention module, and dynamic multimodal fusion module are applied to the reverse denoising process, specifically including: S41: Perform an FFT transform on the current step's pure electrocardiogram estimate and calculate the power spectral density; S42: Divide the frequency domain signal into a low-frequency band of heart rate and a mid-to-high frequency band of morphology, and extract features from each band through a dedicated convolutional branch; S43: The frequency domain fusion features are superimposed onto the noise prediction output with a weight γ=0.05 to correct the denoising mean; S44: Perform parallel grouped convolutions with kernel size=3 and kernel size=7 on the temporal features output by the Transformer; S45: Calculate the attention weight matrix of short-scale features on long-scale features, and fuse them to generate multi-scale temporal features; S46: Superimpose the multi-scale temporal features with a weight β=0.1 onto the original temporal features to update the noise predictor input; S47: Real-time calculation of global cosine similarity between millimeter-wave conditional embedding and electrocardiogram time-domain features; S48: Dynamic fusion weights that map similarity to the [0.3, 0.7] interval using the sigmoid function; S49: Re-weight and fuse multimodal features according to the weights, and update the mechanical domain conditional embedding.

Images