Heart rate detection method and system based on learnable wavelet transform and feature enhancement
By employing learnable wavelet transform and feature enhancement methods, combined with LSTM and Transformer networks, the accuracy and robustness issues of non-contact heart rate monitoring were resolved, achieving high-precision and robust heart rate detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ANHUI UNIV
- Filing Date
- 2026-03-04
- Publication Date
- 2026-06-09
AI Technical Summary
Existing non-contact heart rate monitoring technologies have shortcomings in accuracy and robustness. In particular, they are difficult to accurately separate heartbeat and respiratory signals in situations with low signal-to-noise ratio and variable scenarios. Furthermore, traditional methods rely on complex preprocessing and lack effective long and short time series modeling.
We employ a learnable wavelet transform and feature enhancement method. By adaptively optimizing the learnable wavelet transform of the scale parameter and center frequency, and combining it with a hybrid temporal network of LSTM and Transformer, we perform end-to-end optimization, extract high-resolution time-frequency features, and perform dynamic weighted fusion to output a second-by-second heart rate estimate.
It improves the accuracy and robustness of non-contact heart rate monitoring, enabling accurate tracking of heart rate changes in complex environments, reducing errors, and enhancing sensitivity and anti-interference capabilities to weak heartbeat components.
Smart Images
Figure CN122163178A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the interdisciplinary field of biomedical signal processing and artificial intelligence, specifically to a heart rate detection method and system based on learnable wavelet transform and feature enhancement. Background Technology
[0002] Heart rate is a core physiological indicator reflecting the cardiovascular function and autonomic nervous system regulation of the human body. Traditional heart rate monitoring mainly relies on contact sensors, such as photoplethysmography (PPG) or electrocardiogram (ECG). However, these methods require the subject to wear special devices, which can easily cause skin discomfort and decreased compliance with long-term use, and have significant limitations in scenarios such as sleep, exercise, or infant monitoring. Therefore, the development of non-invasive heart rate monitoring technology that does not require physical contact has become a research hotspot.
[0003] In recent years, non-contact physiological monitoring technology based on frequency-modulated continuous wave millimeter-wave radar has become a research hotspot due to its advantages such as being imperceptible, non-invasive, able to penetrate clothing, and protecting privacy. Its basic principle is that the submillimeter-level displacement of the human chest cavity caused by heartbeat and respiration modulates the phase of the radar echo. This phase signal contains respiratory (0.1–0.5Hz) and heartbeat (0.8–3Hz) components, but the two are heavily coupled and difficult to separate.
[0004] For heart rate estimation of radar phase signals, existing studies mostly employ Continuous Wavelet Transform (CWT) or Convolutional Neural Network (CNN) for feature extraction. While CWT utilizes fixed mother wavelets such as Morlet to perform time-frequency decomposition of the signal and offers superior time-frequency localization capabilities compared to Short-Time Fourier Transform, it still has significant limitations:
[0005] First, as an independent signal processing module, CWT is separated from the subsequent heart rate regression model, making it difficult to optimize overall performance through end-to-end training. Finally, CWT outputs a high-dimensional time-frequency graph containing a large amount of redundant information, which not only increases the computational burden but also easily introduces non-physiological frequency noise interference.
[0006] On the other hand, some studies have attempted to directly input the raw phase signal into a CNN for end-to-end learning. For example, the Chinese invention patent application CN118592921A, "A Method for Detecting Heart Rate and Respiratory Rate in Millimeter-Wave Radar Based on CNN Fusion Features," uses wavelet transform for phase extraction and a trained CNN convolutional neural network to detect heart rate and respiratory rate. However, the shape and scale parameters of the mother wavelet are preset constants, which cannot be dynamically adjusted according to individual heart rate differences or signal-to-noise ratio, resulting in insufficient sensitivity to weak heartbeat components. Secondly, standard CNN convolutional kernels lack prior knowledge of the frequency band characteristics of physiological signals, making it difficult for their local receptive fields to effectively model the periodic oscillation patterns in the 0.8–3Hz heartbeat frequency band. Simultaneously, ordinary convolutional operations have weak frequency selectivity; when the respiratory (0.1–0.5Hz) and heartbeat frequency bands are heavily coupled, respiratory modulation is easily misjudged as the dominant component, causing heart rate estimation errors. Furthermore, CNNs typically rely on extensive manually designed filtering preprocessing or data normalization, weakening their robustness in cross-scenario applications. Summary of the Invention
[0007] The technical problem to be solved by this invention is how to improve the accuracy and robustness of non-contact heart rate monitoring.
[0008] The present invention solves the above-mentioned technical problems through the following technical means: This invention provides a heart rate detection method based on learnable wavelet transform and feature enhancement, comprising the following steps: S1. Acquire the radar phase signal obtained from the millimeter-wave radar and after preprocessing. ; S2, regarding the radar phase signal Perform optimizable scale parameters and center frequency Learnable wavelet transform is used to extract high-resolution time-frequency features covering the heart rate frequency band from 0.8Hz to 3Hz. S3. The high-resolution time-frequency features are decomposed into physiologically relevant frequency bands at multiple scales and dynamically weighted and fused based on the importance weights of the energy of each frequency band, outputting a multi-scale fused feature tensor. S4. Combine the multi-scale fused feature tensor with the original radar phase signal. After projection fusion, the input is given to the LSTM-Transformer hybrid temporal modeling network. The outputs of the LSTM and Transformer are concatenated or added in the time dimension to form a fused feature vector. This vector is then passed through a fully connected regression head to output a second-by-second heart rate estimation sequence. ; S5. During the training phase, use real heart rate labels. With the heart rate estimation sequence Calculate the hybrid loss function and perform end-to-end optimization of the network.
[0009] Furthermore, the radar phase signal acquisition method in step S1 is to use a frequency-modulated continuous wave millimeter-wave radar to transmit a chirp signal, and obtain a 3D radar data cube through multi-channel reception, mixing, and ADC sampling. The preprocessing includes: static clutter filtering, 3D-FFT and beamforming processing of the 3D radar data cube, and extraction of target direction echo; The radar phase signal is obtained by performing phase unwrapping and moving average filtering on the beamformed signal.
[0010] Furthermore, the learnable wavelet transform described in step S2 includes: There are 1D convolutional kernels, each initially set to a Morlet wavelet function, where... For the length of time, The wavelet channel number is expressed as follows:
[0011] in, For scale parameters, ∈[0.8,3.0]Hz, where the center frequency is, the wavelet automatically slides to cover the entire input signal through convolution operation, centered at the zero point of time; all All parameters are treated as learnable parameters and automatically optimized through backpropagation. The output of the learnable wavelet transform is as follows:
[0012] By concatenating the convolutional responses of all channels, a high-resolution time-frequency feature tensor is obtained: .
[0013] Further, step S3 includes the following steps: S31. Using a set of learnable Gabor filters, each set is based on a center frequency. and bandwidth The system controls the input signal and then performs frequency domain decomposition to generate multiple frequency band sub-band features. S32. Calculate the importance weight of each frequency band using a frequency band attention mechanism. The sub-band features are dynamically weighted and fused as follows:
[0014] in This represents the number of Gabor filters. For the first Attention weights for each frequency band For the corresponding frequency band characteristics, It is a learnable projection matrix.
[0015] Furthermore, the hybrid loss function described in step S5 is as follows:
[0016] Wherein, the weight coefficients satisfy , To balance the robustness of L1 loss with the smoothness of L2 loss, the optimizer is RAdam, and the gradient norm clipping threshold is set to 1.0 to ensure stable convergence during training and achieve end-to-end joint optimization.
[0017] This invention also provides a heart rate detection system based on learnable wavelet transform and feature enhancement. The system executes the above-described method during operation and includes the following modules: The data acquisition module is used to acquire radar phase signals obtained from millimeter-wave radar and after preprocessing. ; A learnable wavelet transform module is used to process the radar phase signal. Perform optimizable scale parameters and center frequency Learnable wavelet transform is used to extract high-resolution time-frequency features covering the heart rate frequency band from 0.8Hz to 3Hz. The feature enhancement module is used to decompose the high-resolution time-frequency features into physiologically relevant frequency bands at multiple scales and dynamically weight and fuse them based on the importance weights of the energy of each frequency band, and output a multi-scale fused feature tensor. The network module is used to combine the multi-scale fused feature tensor with the original radar phase signal. After projection fusion, the input is given to the LSTM-Transformer hybrid temporal modeling network. The outputs of the LSTM and Transformer are concatenated or added in the time dimension to form a fused feature vector. This vector is then passed through a fully connected regression head to output a second-by-second heart rate estimation sequence. ; The network training module is used to employ real heart rate labels during the training phase. With the heart rate estimation sequence Calculate the hybrid loss function and perform end-to-end optimization of the network.
[0018] Furthermore, the radar phase signal acquisition method of the data acquisition module is to use a frequency-modulated continuous wave millimeter-wave radar to transmit a chirp signal, and obtain a 3D radar data cube through multi-channel reception, mixing, and ADC sampling. The preprocessing includes: static clutter filtering, 3D-FFT and beamforming processing of the 3D radar data cube, and extraction of target direction echo; The radar phase signal is obtained by performing phase unwrapping and moving average filtering on the beamformed signal.
[0019] Furthermore, the learnable wavelet transform described in the learnable wavelet transform module includes: There are 1D convolutional kernels, each initially set to a Morlet wavelet function, where... For the length of time, The wavelet channel number is expressed as follows:
[0020] in, For scale parameters, ∈[0.8,3.0]Hz, where the center frequency is, the wavelet automatically slides to cover the entire input signal through convolution operation, centered at the zero point of time; all All parameters are treated as learnable parameters and automatically optimized through backpropagation. The output of the learnable wavelet transform is as follows:
[0021] By concatenating the convolutional responses of all channels, a high-resolution time-frequency feature tensor is obtained: .
[0022] Furthermore, the feature enhancement module includes the following units: Multi-band sub-band feature units are used to utilize a set of learnable Gabor filters, each set consisting of a center frequency. and bandwidth The system controls the input signal and then performs frequency domain decomposition to generate multiple frequency band sub-band features. The sub-band feature fusion unit is used to calculate the importance weight of each frequency band through a frequency band attention mechanism. The sub-band features are dynamically weighted and fused as follows:
[0023] in This represents the number of Gabor filters. For the first Attention weights for each frequency band For the corresponding frequency band characteristics, It is a learnable projection matrix.
[0024] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the above-described method.
[0025] The advantages of this invention are: (1) This invention employs adaptively optimizeable scale parameters and center frequency The learnable wavelet transform is used to extract high-resolution time-frequency features covering the heart rate frequency band from 0.8Hz to 3Hz; and a learnable filter bank and frequency band attention mechanism are introduced to refine the modeling and dynamic weighting of physiologically relevant frequency bands, thereby achieving super-resolution-level frequency domain feature enhancement, improving the sensitivity to weak heartbeat components, and thus improving the accuracy of non-contact heartbeat detection. (2) This invention improves the robustness of the network by introducing a hybrid temporal network composed of LSTM and Transformer. LSTM captures the periodic local dynamics of heartbeat, while Transformer models long-range physiological dependence. Attached Figure Description
[0026] Figure 1 This is a schematic diagram of the heart rate detection method based on learnable wavelet transform and feature enhancement according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the actual test results of the present invention on the IWR1642 radar; Figure 3 This is a comparison chart of heart rate estimation results between the embodiments of the present invention and traditional signal processing methods; Figure 4 This is a performance comparison chart between the embodiments of the present invention and a single-structure neural network model. Detailed Implementation
[0027] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0028] Example 1 The purpose of this embodiment is to address the shortcomings of existing non-contact heart rate detection methods, such as insufficient accuracy in low signal-to-noise ratio environments, weak generalization ability, reliance on complex preprocessing, and lack of effective long and short-term time series modeling. It aims to provide a high-precision, robust, end-to-end, and data-normalization-free non-contact heart rate detection method. This method can accurately estimate the heart rate per second directly from the raw millimeter-wave radar phase signal, making it suitable for non-intrusive health monitoring in real-world complex scenarios. This embodiment provides a heart rate detection method based on learnable wavelet transform and feature enhancement, with the specific implementation process as follows: Figure 1 As shown, it includes the following steps: S1. Acquire the radar phase signal obtained from the millimeter-wave radar and after preprocessing. The radar phase signal is acquired by transmitting a chirp signal using a frequency-modulated continuous wave millimeter-wave radar. Multiple receiving antennas synchronously receive the echo signal reflected from the human body, and the signal is down-converted to an intermediate frequency signal through a mixer circuit. The intermediate frequency signal is then sampled by a high-speed ADC to generate a three-dimensional radar data cube (containing distance, time, and angle dimensions). Several initial chirp frames are removed to suppress the sensor's transient response, and static clutter filtering is performed to reduce environmental noise. The preprocessing includes: static clutter filtering of the 3D radar data cube; 3D-FFT operation on the denoised 3D data; determining the target azimuth angle by combining coherent accumulation; beamforming using the conjugate steering vector weighting method to enhance the target direction signal and suppress interference from other directions; and subsequently, phase unwrapping and moving average filtering of the beamformed signal sequence to obtain the vital signs micro-motion radar phase signal for subsequent analysis. .
[0029] S2, regarding the radar phase signal Adaptive optimization of scale parameters and center frequency The learnable wavelet transform extracts high-resolution time-frequency features covering the heart rate frequency band from 0.8Hz to 3Hz; the learnable wavelet transform includes: There are 1D convolutional kernels, each initially set to a Morlet wavelet function, where... For the length of time, The wavelet channel number is expressed as follows:
[0030] in, For scale parameters, ∈[0.8,3.0]Hz, where the center frequency is, the wavelet is centered at the zero point of time (i.e. ); automatically slides to cover the entire input signal through convolution operations; all All parameters are treated as learnable parameters and automatically optimized through backpropagation. The output of the learnable wavelet transform is as follows:
[0031] During model training, automatic optimization enables each convolutional kernel to adaptively focus on the physiological frequency band that is most discriminative to the current sample.
[0032] By concatenating the convolutional responses of all channels, a high-resolution time-frequency feature tensor is obtained: .
[0033] In this embodiment, a high-resolution time-frequency feature covering the 0.8Hz to 3Hz heart rate frequency band is generated, with an output dimension of B×T×D1; at the same time, the linear projection of the original signal is retained as a supplementary feature, with an output dimension of B×T×D3. In this embodiment, B is 32, T is 32, and D1 and D3 are both 64.
[0034] S3. The high-resolution time-frequency features are obtained by multi-scale decomposition of physiologically relevant frequency bands and dynamic weighted fusion based on the importance weights of the energy of each frequency band, outputting a multi-scale fused feature tensor; the specific implementation includes the following steps: S31. Using a set of learnable Gabor filters, each set is based on a center frequency. and bandwidth The system controls the input signal and then performs frequency domain decomposition to generate multiple frequency band sub-band features. S32. Calculate the importance weight of each frequency band using a frequency band attention mechanism. The sub-band features are dynamically weighted and fused as follows:
[0035] in This represents the number of Gabor filters. For the first Attention weights for each frequency band For the corresponding frequency band characteristics, It is a learnable projection matrix.
[0036] In this embodiment, the initial center frequency is set in the range of 0.8–3Hz, and the bandwidth is adjustable; by dynamically weighting the energy of each frequency band through frequency band attention, the refined modeling and enhancement of physiologically relevant frequency bands are achieved, and the output dimension is B×T×D2; the value of D2 is 64.
[0037] S4. Combine the multi-scale fused feature tensor with the original radar phase signal. After projection fusion, the input is given to an LSTM-Transformer hybrid temporal modeling network. The LSTM sub-network extracts short-term dependencies within the heartbeat cycle, and its hidden state evolution follows the standard LSTM recursive formula. The Transformer encoder sub-network captures long-range contextual dependencies across multiple heartbeat cycles through a multi-head self-attention mechanism. The outputs of the LSTM and Transformer are concatenated or added in the time dimension to form a fused feature vector, which is then passed through a fully connected regression head to output a second-by-second heart rate estimation sequence. The unit is bpm.
[0038] S5. During the training phase, use real heart rate labels. With the heart rate estimation sequence Calculate the hybrid loss function and perform end-to-end optimization of the network.
[0039] The hybrid loss function is as follows:
[0040] Wherein, the weight coefficients satisfy , To balance the robustness of L1 loss with the smoothness of L2 loss, the optimizer RAdam was chosen, with a gradient norm clipping threshold of 1.0 to ensure stable convergence during training and achieve end-to-end joint optimization. In this embodiment, the model was trained on the public Monitoring Dataset, with training samples including synchronously acquired radar phase signals and reference heart rates. The input sequence length was 1200 points, corresponding to a duration of 60 seconds. The dataset contained 440 original samples, which were split using a sliding window method with a window length of 10 seconds and a sliding step size of 1 second, generating approximately 538,560 training samples. The batch size was set to 32, and the loss function was a weighted hybrid loss of L1 and L2.
[0041] L1 represents the mean absolute error (MAE), and L2 represents the mean squared error (MSE). The weights are set to balance the model's robustness to outliers and the smoothness of predictions. The optimizer used is RAdam, with an initial learning rate of 0.001. Gradient clipping (norm upper limit 1.0) and cosine annealing learning rate scheduling strategies are employed, and the training epochs are 300.
[0042] To verify the effectiveness of this method, an experiment was conducted. The experimental environment was an indoor office desk, with the subject facing the radar sensor at a distance of approximately 1 meter, wearing a pulse oximeter to simultaneously record their real heart rate. The radar sensor was connected to a computer via a data acquisition card, running data acquisition software to achieve simultaneous acquisition and processing of radar and physiological signals. All data acquisition complied with ethical guidelines, and the subject provided informed consent. The entire process required no contact with the human body, achieving truly contactless monitoring. The experiment was tested on the TI AWR1642 millimeter-wave radar platform. Figure 2 As shown, this serves as a supplement to the test set and is used for testing and verification. In the experiment, the subjects sat approximately 1 meter in front of the radar, and a fingertip pulse oximeter was used as the reference heart rate source, continuously collecting data for 45 seconds. In the figure, the red curve represents the heart rate predicted by the radar system, and the blue curve represents the actual value measured by the pulse oximeter. The results show that this method can accurately track the trend of heart rate changes, maintaining good consistency under normal breathing and slight body movement conditions, with small errors. The mean absolute error (MAE) is less than 3 bpm, indicating that the model has strong robustness and practicality.
[0043] like Figure 3 As shown, the heart rate detection results of this method were compared with those of traditional signal processing methods FFT and CWT on volunteers of two different genders. The "true value" is the reference heart rate. It can be seen that the CWT method exhibits significant distortion or jumps during periods of intense physical activity (e.g., at the 30th second for tester 1 and the 10th second for tester 2), while the prediction curve of this method consistently closely approximates the true value with smooth fluctuations, significantly improving its anti-interference capability.
[0044] like Figure 4 As shown, the performance of our method is compared with that of single-structure neural network models (pure LSTM and pure Transformer). Under the same test conditions, our method (red) exhibits the best tracking accuracy and stability; pure LSTM (blue) performs reasonably well in the long term, but lags behind in responding to instantaneous changes; pure Transformer (purple), while able to capture global patterns, suffers from overfitting at local abrupt changes, leading to oscillations in predicted values. This indicates that the hybrid structure of LSTM and Transformer effectively balances local fine-grained modeling with global context awareness. This invention enhances model stability by achieving dual-stream information fusion through residual connections.
[0045] This method constructs a high-precision and robust non-contact heart rate detection method by integrating continuous wavelet transform, super-resolution feature enhancement, and an LSTM-Transformer hybrid network. The experimental results show that this method outperforms traditional spectral analysis and single neural network models in real-world environments, demonstrating good potential for clinical application.
[0046] Example 2 It should be further explained that, based on the same inventive concept, this embodiment provides a heart rate detection system based on learnable wavelet transform and feature enhancement. When the system runs, it executes the method described in Embodiment 1, including the following modules: The data acquisition module is used to acquire radar phase signals obtained from millimeter-wave radar and after preprocessing. ; A learnable wavelet transform module is used to process the radar phase signal. Perform optimizable scale parameters and center frequency Learnable wavelet transform is used to extract high-resolution time-frequency features covering the heart rate frequency band from 0.8Hz to 3Hz. The feature enhancement module is used to decompose the high-resolution time-frequency features into physiologically relevant frequency bands at multiple scales and dynamically weight and fuse them based on the importance weights of the energy of each frequency band, and output a multi-scale fused feature tensor. The network module is used to combine the multi-scale fused feature tensor with the original radar phase signal. After projection fusion, the input is given to the LSTM-Transformer hybrid temporal modeling network. The outputs of the LSTM and Transformer are concatenated or added in the time dimension to form a fused feature vector. This vector is then passed through a fully connected regression head to output a second-by-second heart rate estimation sequence. ; The network training module is used to employ real heart rate labels during the training phase. With the heart rate estimation sequence Calculate the hybrid loss function and perform end-to-end optimization of the network.
[0047] The radar phase signal acquisition method of the data acquisition module is to use frequency-modulated continuous wave millimeter-wave radar to transmit chirp signals, and obtain a 3D radar data cube through multi-channel reception, mixing, and ADC sampling. The preprocessing includes: static clutter filtering, 3D-FFT and beamforming processing of the 3D radar data cube, and extraction of target direction echo; The radar phase signal is obtained by performing phase unwrapping and moving average filtering on the beamformed signal.
[0048] The learnable wavelet transform described in the learnable wavelet transform module includes: There are 1D convolutional kernels, each initially set to a Morlet wavelet function, where... For the length of time, The wavelet channel number is expressed as follows:
[0049] in, For scale parameters, ∈[0.8,3.0]Hz, where the center frequency is, the wavelet automatically slides to cover the entire input signal through convolution operation, centered at the zero point of time; all All parameters are treated as learnable parameters and automatically optimized through backpropagation. The output of the learnable wavelet transform is as follows:
[0050] By concatenating the convolutional responses of all channels, a high-resolution time-frequency feature tensor is obtained: .
[0051] The feature enhancement module includes the following units: Multi-band sub-band feature units are used to utilize a set of learnable Gabor filters, each set consisting of a center frequency. and bandwidth The system controls the input signal and then performs frequency domain decomposition to generate multiple frequency band sub-band features. The sub-band feature fusion unit is used to calculate the importance weight of each frequency band through a frequency band attention mechanism. The sub-band features are dynamically weighted and fused as follows:
[0052] in This represents the number of Gabor filters. For the first Attention weights for each frequency band For the corresponding frequency band characteristics, It is a learnable projection matrix.
[0053] This embodiment also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in Embodiment 1.
[0054] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A heart rate detection method based on learnable wavelet transform and feature enhancement, characterized in that, Includes the following steps: S1. Acquire the radar phase signal obtained from the millimeter-wave radar and after preprocessing. ; S2, regarding the radar phase signal Adaptive optimization of scale parameters and center frequency Learnable wavelet transform is used to extract high-resolution time-frequency features covering the heart rate frequency band from 0.8Hz to 3Hz. S3. The high-resolution time-frequency features are decomposed into physiologically relevant frequency bands at multiple scales and dynamically weighted and fused based on the importance weights of the energy of each frequency band, outputting a multi-scale fused feature tensor. S4. Combine the multi-scale fused feature tensor with the original radar phase signal. After projection fusion, the input is given to the LSTM-Transformer hybrid temporal modeling network. The outputs of the LSTM and Transformer are concatenated or added in the time dimension to form a fused feature vector. This vector is then passed through a fully connected regression head to output a second-by-second heart rate estimation sequence. ; S5. During the training phase, use real heart rate labels. With the heart rate estimation sequence Calculate the hybrid loss function and perform end-to-end optimization of the network.
2. The heart rate detection method based on learnable wavelet transform and feature enhancement according to claim 1, characterized in that, The radar phase signal acquisition method in step S1 is to use a frequency-modulated continuous wave millimeter-wave radar to transmit a chirp signal, and obtain a 3D radar data cube through multi-channel reception, mixing, and ADC sampling. The preprocessing includes: static clutter filtering, 3D-FFT and beamforming processing of the 3D radar data cube, and extraction of target direction echo; The radar phase signal is obtained by performing phase unwrapping and moving average filtering on the beamformed signal.
3. The heart rate detection method based on learnable wavelet transform and feature enhancement according to claim 1, characterized in that, The learnable wavelet transform mentioned in step S2 includes: There are 1D convolutional kernels, each initially set to a Morlet wavelet function, where... For the length of time, The wavelet channel number is expressed as follows: in, For scale parameters, ∈[0.8,3.0]Hz, where the center frequency is, the wavelet automatically slides to cover the entire input signal through convolution operation, centered at the zero point of time; all All parameters are treated as learnable parameters and automatically optimized through backpropagation. The output of the learnable wavelet transform is as follows: By concatenating the convolutional responses of all channels, a high-resolution time-frequency feature tensor is obtained: 。 4. The heart rate detection method based on learnable wavelet transform and feature enhancement according to claim 1, characterized in that, Step S3 includes the following steps: S31. Using a set of learnable Gabor filters, each set is based on a center frequency. and bandwidth The system controls the input signal and then performs frequency domain decomposition to generate multiple frequency band sub-band features. S32. Calculate the importance weight of each frequency band using a frequency band attention mechanism. The sub-band features are dynamically weighted and fused as follows: in This represents the number of Gabor filters. For the first Attention weights for each frequency band For the corresponding frequency band characteristics, It is a learnable projection matrix.
5. The heart rate detection method based on learnable wavelet transform and feature enhancement according to claim 1, characterized in that, The hybrid loss function described in step S5 is as follows: Wherein, the weight coefficients satisfy , To balance the robustness of L1 loss with the smoothness of L2 loss, the optimizer is RAdam, and the gradient norm clipping threshold is set to 1.0 to ensure stable convergence during training and achieve end-to-end joint optimization.
6. A heart rate detection system based on learnable wavelet transform and feature enhancement, characterized in that, Includes the following modules: The data acquisition module is used to acquire radar phase signals obtained from millimeter-wave radar and after preprocessing. ; A learnable wavelet transform module is used to process the radar phase signal. Perform optimizable scale parameters and center frequency Learnable wavelet transform is used to extract high-resolution time-frequency features covering the heart rate frequency band from 0.8Hz to 3Hz. The feature enhancement module is used to decompose the high-resolution time-frequency features into physiologically relevant frequency bands at multiple scales and dynamically weight and fuse them based on the importance weights of the energy of each frequency band, and output a multi-scale fused feature tensor. The network module is used to combine the multi-scale fused feature tensor with the original radar phase signal. After projection fusion, the input is given to the LSTM-Transformer hybrid temporal modeling network. The outputs of the LSTM and Transformer are concatenated or added in the time dimension to form a fused feature vector. This vector is then passed through a fully connected regression head to output a second-by-second heart rate estimation sequence. ; The network training module is used to employ real heart rate labels during the training phase. With the heart rate estimation sequence Calculate the hybrid loss function and perform end-to-end optimization of the network.
7. The heart rate detection system based on learnable wavelet transform and feature enhancement according to claim 6, characterized in that, The radar phase signal acquisition method of the data acquisition module is to use frequency-modulated continuous wave millimeter-wave radar to transmit chirp signals, and obtain a 3D radar data cube through multi-channel reception, mixing, and ADC sampling. The preprocessing includes: static clutter filtering, 3D-FFT and beamforming processing of the 3D radar data cube, and extraction of target direction echo; The radar phase signal is obtained by performing phase unwrapping and moving average filtering on the beamformed signal.
8. The heart rate detection system based on learnable wavelet transform and feature enhancement according to claim 6, characterized in that, The learnable wavelet transform described in the learnable wavelet transform module includes: There are 1D convolutional kernels, each initially set to a Morlet wavelet function, where... For the length of time, The wavelet channel number is expressed as follows: in, For scale parameters, ∈[0.8,3.0]Hz, where the center frequency is, the wavelet automatically slides to cover the entire input signal through convolution operation, centered at the zero point of time; all All parameters are treated as learnable parameters and automatically optimized through backpropagation. The output of the learnable wavelet transform is as follows: By concatenating the convolutional responses of all channels, a high-resolution time-frequency feature tensor is obtained: 。 9. The heart rate detection system based on learnable wavelet transform and feature enhancement according to claim 6, characterized in that, The feature enhancement module includes the following units: Multi-band sub-band feature units are used to utilize a set of learnable Gabor filters, each set consisting of a center frequency. and bandwidth The system controls the input signal and then performs frequency domain decomposition to generate multiple frequency band sub-band features. The sub-band feature fusion unit is used to calculate the importance weight of each frequency band through a frequency band attention mechanism. The sub-band features are dynamically weighted and fused as follows: in This represents the number of Gabor filters. For the first Attention weights for each frequency band For the corresponding frequency band characteristics, It is a learnable projection matrix.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-5.