BCG signal data processing method based on TimeGAN model

Abstract

The invention relates to the technical field of biomedical signal processing, in particular to a BCG signal data processing method based on a TimeGAN model, and the method comprises the steps: obtaining an original BCG signal data set X, and carrying out the preprocessing of the original BCG signal data set X, and obtaining time-sequence formatted BCG data; the preprocessing comprises denoising, interpolation completion, normalization and time alignment; the time sequence formatted BCG data are input into the trained TimeGAN model, and an enhanced BCG signal Xfault is generated; and carrying out noise suppression processing on the enhanced BCG signal Xfault to obtain a de-noised enhanced BCG signal Xdeniosed, and carrying out noise suppression processing on the enhanced BCG signal Xfault to obtain a de-noised enhanced BCG signal Xdeniosed. According to the invention, the TimeGAN model is utilized to generate the enhanced BCG signal, and the frequency domain denoising technology is combined, so that the signal quality and the time sequence consistency are improved. The method effectively complements missing data, inhibits noise, enhances data availability, and improves the accuracy and reliability of health monitoring and medical diagnosis.

Description

Technical Field

[0001] The present invention relates to the technical field of biomedical signal processing, and in particular to a method for processing BCG signal data based on the TimeGAN model. Background Art

[0002] Ballistocardiography (BCG) is a non-invasive cardiovascular physiological signal that can reflect the weak body movements caused by heart beats. Due to its simple acquisition method and non-invasiveness to the human body, BCG signals have broad application prospects in the fields of remote health monitoring, intelligent medical devices, and physiological signal analysis. However, the amplitude of BCG signals is small and is easily affected by factors such as environmental noise, individual differences, and sensor drift, resulting in low signal quality and thus affecting subsequent signal analysis and applications.

[0003] Currently, the processing methods for BCG signals mainly include techniques such as filtering and denoising, interpolation and completion, and signal enhancement. However, traditional filtering methods usually rely on fixed frequency band settings and are difficult to effectively distinguish useful signals from noise, which may lead to signal loss or the introduction of artifacts. In addition, due to the strong temporal dependence of BCG signals, existing deep learning-based signal generation methods (such as variational autoencoders VAE, generative adversarial networks GAN, etc.) are difficult to effectively capture the dynamic temporal characteristics of signals, resulting in insufficient authenticity of the generated signals and unable to meet the requirements of medical applications. Summary of the Invention

[0004] In view of the above-mentioned disadvantages and deficiencies of the prior art, the present invention provides a method for processing BCG signal data based on the TimeGAN model.

[0005] To achieve the above object, the main technical solutions adopted by the present invention include:

[0006] An embodiment of the present invention provides a method for processing BCG signal data based on the TimeGAN model, including:

[0007] S1. Obtain the original BCG signal dataset X and perform preprocessing on it to obtain time-series formatted BCG data;

[0008] The preprocessing includes denoising, interpolation and completion, normalization, and time alignment;

[0009] S2. Input the time-series formatted BCG data into the trained TimeGAN model to generate enhanced BCG signals X fake ;

[0010] S3. Perform noise suppression processing on the enhanced BCG signals X fake to obtain denoised enhanced BCG signals Xdenoised 。

[0011] Preferably, step S3 specifically includes:

[0012] S31. Use the fast Fourier transform to convert the enhanced BCG signal X fake to the frequency domain to obtain the spectral component F fake ;

[0013] S32. Calculate the frequency attention weight W f and apply it to the spectral component F fake to suppress high-frequency noise and obtain the optimized frequency-domain signal F denoised ;

[0014] S33. Use the inverse Fourier transform to convert the frequency-domain signal F denoised back to the time domain to obtain the denoised enhanced BCG signal X denoised 。

[0015] Preferably, before step S1, it further includes:

[0016] S0. Use the training data set to train the pre-obtained TimeGAN model to obtain the trained TimeGAN model;

[0017] The training data set includes: time-series formatted BCG data for training the TimeGAN model in the historical time period;

[0018] The TimeGAN model includes:

[0019] A time-series encoder for receiving the time-series formatted BCG data input into the TimeGAN model, extracting time-series features, and converting them into a latent space representation H real ;

[0020] A generator for generating a latent representation H from random noise Z fake ;

[0021] A time-series decoder for restoring the latent space representation H real and the latent representation H fake back to the time-domain BCG signal, where the data restored from the latent representation H fake is the generated BCG data;

[0022] A discriminator for distinguishing the difference between the latent space representation H real and the latent representation H fake as well as between the time-series formatted BCG data and the generated BCG data.

[0023] Preferably,

[0024] The timing encoder uses two layers of gated recurrent units and one layer of fully connected layers for feature extraction. The first layer of gated recurrent units extracts basic timing characteristics, the second layer of gated recurrent units captures complex timing dependencies, and the fully connected layer is used to convert the timing features into a latent space representation.

[0025] Preferably,

[0026] Among them, the frequency attention weight W is calculated using formula (1) f ;

[0027] The formula (1) is:

[0028]

[0029] Among them, β is a preset steepness value;

[0030] T is the threshold for setting noise suppression;

[0031] |F fake | is the amplitude information of the BCG signal in the frequency domain.

[0032] Preferably, the TimeGAN model uses a self-supervised learning loss function during training;

[0033] The self-supervised learning loss function is:

[0034] L TimeGAN =γ1L rec +γ2L gen +γ3L adv +γ4L sup ;

[0035] L rec is the reconstruction loss; L gen is the generation loss; L adv is the adversarial loss; L sup is the temporal consistency loss; γ1 is the first hyperparameter; γ2 is the second hyperparameter; γ3 is the third hyperparameter; γ4 is the fourth hyperparameter.

[0036] Preferably, among them,

[0037]

[0038] Among them, is the reconstructed data generated by the encoder-decoder;

[0039] E[] represents the expected sum of all data in [];

[0040] L gen =E[log D gen (G(Z))];

[0041] Among them, G(Z) is the data generated by the generator; D gen is the discriminator's judgment on the synthetic data;

[0042] L adv = E[log D(X) + log(1 - D(G(Z)))];

[0043] Among them, D(X) is the discriminator's score for the real data; D(G(Z)) is the score for the generated data;

[0044]

[0045] Among them, h t is the hidden state at time step t; is the hidden state obtained through the time series encoder - time series decoder.

[0046] Preferably, when training the TimeGAN model, the random batch technique is adopted, and each batch of data includes 64 - 128 time series formatted BCG data.

[0047] Preferably, during the training process of the TimeGAN model, a weighted adversarial loss function is adopted to enhance the temporal correlation between the generated BCG data and the time series formatted BCG data;

[0048] The weighted adversarial loss function is:

[0049] L adv = E[log D(X real )] + δ·E[log(1 - D(X fake ))];

[0050] L adv is the weighted adversarial loss function;

[0051] D(·) is the output probability of the discriminator;

[0052] X real is the real BCG signal;

[0053] X fake is the generated BCG signal;

[0054] δ is a preset dynamic weighting factor;

[0055] E[] represents the expected sum of all data in [].

[0056] Preferably, during the training process of the TimeGAN model, the Adam optimizer is used to optimize the parameters of the generator, discriminator, temporal encoder, and temporal decoder to improve the quality and temporal consistency of the generated signals.

[0057] The beneficial effects of the present invention are as follows:

[0058] A BCG signal data processing method based on the TimeGAN model of the present invention can more effectively retain the temporal characteristics of the BCG signal, suppress high-frequency noise, improve the signal quality, enhance the authenticity and usability of the BCG signal, and thus enhance the application value of the BCG signal in remote health monitoring and intelligent medical devices, because it uses the TimeGAN model to generate enhanced BCG signals and combines frequency domain denoising technology, compared with the prior art.

[0059] A BCG signal data processing method based on the TimeGAN model of the present invention can more accurately suppress high-frequency noise and maintain the key temporal information of the signal, achieving the purpose of optimizing the signal quality and reducing noise interference, and improving the applicability of the BCG signal in health monitoring and medical analysis, because the fast Fourier transform (FFT) and frequency attention mechanism are introduced during the signal denoising process, compared with the prior art.

[0060] A BCG signal data processing method based on the TimeGAN model of the present invention can effectively reduce the problem of data quality degradation caused by sensor drift or signal loss, achieving the effect of improving the integrity and consistency of BCG signal data and making it more suitable for subsequent analysis and application, because preprocessing techniques such as normalization, interpolation completion, and time alignment are used, compared with the prior art.

[0061] A BCG signal data processing method based on the TimeGAN model of the present invention can more accurately learn the temporal characteristics of the BCG signal and generate high-quality enhanced signals that conform to the true distribution, achieving the effect of improving data integrity and enhancing data diversity, and contributing to enhancing the reliability and stability of the BCG signal, because the TimeGAN model combining a temporal encoder, a generator, a temporal decoder, and a discriminator is used, compared with the prior art. Description of the Drawings

[0062] Figure 1 It is a flowchart of a BCG signal data processing method based on the TimeGAN model of the present invention. Detailed Embodiments

[0063] To better explain the present invention for easy understanding, the present invention will be described in detail below with reference to the drawings and through specific embodiments.

[0064] To better understand the above technical solution, the exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a clearer and more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.

[0065] Embodiment 1

[0066] See Figure 1 , this embodiment provides a BCG signal data processing method based on the TimeGAN model, including:

[0067] S1. Obtain the original BCG signal dataset X and preprocess it to obtain the time-series formatted BCG data;

[0068] The preprocessing includes denoising, interpolation and completion, normalization, and time alignment;

[0069] In step S1, preprocessing techniques such as denoising, interpolation and completion, normalization, and time alignment are used to effectively solve the common problems of data loss, noise interference, and time asynchronization in the acquisition process of BCG signals. For example, in remote health monitoring devices, due to unstable sensor contact or individual movement, the BCG signal may have missing values or large noise. Through the interpolation and completion technique, the missing waveform data can be restored; normalization makes the amplitude ranges of BCG signals from different individuals unified, facilitating model learning, thereby improving the integrity and comparability of signal data.

[0070] S2. Input the time-series formatted BCG data into the trained TimeGAN model to generate the enhanced BCG signal X fake ;

[0071] In step S2, the TimeGAN model is used to generate and enhance the BCG signal. Compared with traditional GANs or variational autoencoders (VAEs), TimeGAN combines a time-series encoder and decoder, which can better learn the temporal features of the BCG signal and make the generated enhanced signal more conform to the real physiological waveform. For example, in medical research, there is less BCG data under certain pathological conditions, which is difficult to support effective model training. However, TimeGAN can generate additional realistic samples to enrich the dataset, improve the generalization ability of the model, and thus improve the performance of health monitoring algorithms.

[0072] S3. Perform noise suppression processing on the enhanced BCG signal X fake to obtain the denoised enhanced BCG signal X denoised .

[0073] In step S3, by using the fast Fourier transform (FFT) combined with the frequency attention mechanism, while effectively removing high-frequency noise, the key feature information of the BCG signal is retained. For example, in practical applications, the BCG signal may be interfered by the electrocardiogram (ECG) signal, electromyogram (EMG) signal, or environmental vibration, and these interferences are mainly distributed in the high-frequency region. In step S31, the signal is transformed into the frequency domain through FFT, making the noise easier to identify; the frequency attention weight W calculated in step S32 f can dynamically adjust the signal processing strategy, only attenuating high-frequency noise without excessive loss of effective signals. For example, in a sleep monitoring device, this method can effectively reduce signal artifacts caused by factors such as breathing and body movement, thereby improving the quality of the BCG signal and making the cardiovascular health assessment more accurate.

[0074] In this embodiment, the specific steps of S3 include:

[0075] S31. Use the fast Fourier transform to transform the enhanced BCG signal X fake into the frequency domain to obtain the spectral component F fake ;

[0076] In this embodiment, in the time domain, noise and effective signals are often mixed together and difficult to distinguish; while in the frequency domain, different components are distributed in different frequency regions, making noise processing more accurate. Physiological signals (such as heart rate information) are generally concentrated in the low-frequency band, while interferences (such as mechanical vibration, electronic device noise) are often in the high-frequency band. After being transformed into the frequency domain, specific frequency ranges can be optimized without affecting the overall signal quality.

[0077] S32. Calculate the frequency attention weight W f and apply it to the spectral component F fake to suppress high-frequency noise and obtain the optimized frequency domain signal F denoised ;

[0078] Among them, the frequency attention weight W is calculated using formula (1) f ;

[0079] The formula (1) is:

[0080]

[0081] Among them, β is a preset steepness value; T is the threshold for setting noise suppression; |F fake | is the amplitude information of the BCG signal in the frequency domain.

[0082] Traditional filtering methods (such as low-pass filtering) may accidentally damage some useful signals, while the frequency attention mechanism can dynamically adjust the denoising intensity according to the signal amplitude size to reduce signal loss. β controls the steepness of the filtering, making the attenuation of high-frequency signals not sudden but smooth, which can reduce the edge effect of the signal and avoid losing important high-frequency information (such as heart sounds, etc.). The values of β and T can be adjusted according to different application scenarios to make the algorithm applicable to different BCG signal noise situations. For example: Quiet environment (such as sleep monitoring): T can be set higher to reduce the impact on tiny high-frequency components. Complex environment (such as exercise state): T is set lower to enhance noise suppression.

[0083] S33. Use the inverse Fourier transform to convert the frequency-domain signal F denoised back to the time domain to obtain the enhanced BCG signal X after denoising denoised .

[0084] Traditional noise filtering methods, such as fixed-threshold filtering or wavelet transform filtering, may lead to the loss of key physiological characteristics of the signal. In this method, the inverse Fourier transform (IFFT) (step S33) is used to convert the optimized signal back to the time domain to ensure that the original waveform of the signal is retained. For example, in heart rate variability (HRV) analysis, the peak and rhythm characteristics of the BCG signal are crucial. Using this method can retain these key information while denoising, making the subsequent heart rate analysis more accurate and reliable.

[0085] In this embodiment, before S1, it further includes:

[0086] S0. Use the training dataset to train the pre-obtained TimeGAN model to obtain the trained TimeGAN model;

[0087] The training dataset includes: the time-series formatted BCG data for training the TimeGAN model in the historical time period;

[0088] The TimeGAN model includes:

[0089] A time-series encoder for receiving the time-series formatted BCG data input into the TimeGAN model, extracting time-series features, and converting them into a latent space representation H real ; Specifically, the time-series encoder uses two layers of gated recurrent units and one layer of fully connected layer for feature extraction. The first layer of gated recurrent unit extracts basic time-series characteristics, the second layer of gated recurrent unit captures complex time-series dependencies, and the fully connected layer is used to convert the time-series features into a latent space representation.

[0090] A generator for generating a latent representation H from the random noise Z fake ;

[0091] A temporal decoder for converting the latent space representation H real and the latent representation H fake back into the temporal BCG signal, where the data restored from the latent representation H fake is the generated BCG data;

[0092] A discriminator for distinguishing the difference between the latent space representation H real and the latent representation H fake as well as the temporally formatted BCG data and the generated BCG data.

[0093] In the TimeGAN model, the temporal encoder adopts a two-layer gated recurrent unit (GRU) + fully connected layer for feature extraction: The first layer of GRU extracts basic temporal features such as the heartbeat period and respiratory rhythm. The second layer of GRU captures more complex temporal dependencies such as heart rate variability (HRV) and the modulation effect of respiration on the heartbeat. The fully connected layer is used to convert it into the latent space representation (Hreal), ensuring that the data representation in the latent space is more compact and smooth, which helps with subsequent generation and decoding. It ensures the temporal consistency of the generated data, preventing sudden changes or unreasonable time series patterns. It effectively retains short-term and long-term dependencies and is suitable for physiological signal analysis tasks that require high temporal resolution, such as cardiovascular disease detection and sleep monitoring. It is more suitable for physiological signals with strong temporal dependencies like BCG compared to traditional GANs.

[0094] In the practical application of this embodiment, the TimeGAN model adopts a self-supervised learning loss function during training;

[0095] The self-supervised learning loss function is:

[0096] L TimeGAN = γ1L rec + γ2L gen + γ3L adv + γ4L sup ;

[0097] L rec is the reconstruction loss; L gen is the generation loss; L adv is the adversarial loss; L sup is the temporal consistency loss; γ1 is the first hyperparameter; γ2 is the second hyperparameter; γ3 is the third hyperparameter; γ4 is the fourth hyperparameter.

[0098] Among them,

[0099] Among them, is the reconstructed data generated by the encoder-decoder;

[0100] $E[\ ]$ represents the expected sum over all data in $[\ ]$;

[0101] L gen $= E[\log D gen (G(Z))]$;

[0102] where $G(Z)$ is the data generated by the generator; $D gen is the discriminator's judgment on the synthetic data;

[0103] L adv $= E[\log D(X)+\log(1 - D(G(Z)))]$;

[0104] where $D(X)$ is the discriminator's score on the real data; $D(G(Z))$ is the score on the generated data;

[0105]

[0106] where $h t is the hidden state at time step $t$; is the hidden state obtained through the temporal encoder - temporal decoder.

[0107] In this embodiment, when training the TimeGAN model, the stochastic batch technique is adopted, and each batch of data includes 64 - 128 temporally formatted BCG data. After adopting the stochastic batch technique in this embodiment, the data gradients of only 64 - 128 samples are calculated each time, resulting in less computational overhead and improving the training efficiency. The BCG signal has strong temporal correlation, and the stochastic batch technique can dynamically establish connections between data in different time periods, enabling TimeGAN to better learn long - term dependencies.

[0108] During the training process of the TimeGAN model, a weighted adversarial loss function is adopted to enhance the temporal correlation of the generated BCG data and the temporally formatted BCG data;

[0109] The weighted adversarial loss function is:

[0110] L adv $= E[\log D(X real )]+\delta\cdot E[\log(1 - D(X fake )]$;

[0111] L adv is the weighted adversarial loss function;

[0112] $D(\cdot)$ is the output probability of the discriminator;

[0113] X real is the real BCG signal;

[0114] Xfake is the generated BCG signal;

[0115] δ is a preset dynamic weighting factor;

[0116] E[] represents the expected sum of all data in [].

[0117] The loss function of the ordinary GAN only focuses on the similarity of data distribution and does not optimize for temporal correlation. In this embodiment, the TimeGAN model adopts a weighted adversarial loss function during the training process to adjust the loss terms, making the data generated by the TimeGAN model more realistic at the time series level.

[0118] During the training process of the TimeGAN model described in this embodiment, the Adam optimizer is used to optimize the parameters of the generator, discriminator, temporal encoder, and temporal decoder to improve the quality and temporal consistency of the generated signal.

[0119] Embodiment 2

[0120] The technical solution of this Embodiment 2 relates to the field of medical signal processing, and specifically relates to a method for BCG (Ballistocardiography) signal data enhancement and denoising based on TimeGAN, which is applicable to application scenarios such as remote health monitoring, intelligent medical devices, and cardiovascular health assessment.

[0121] This method uses TimeGAN (Time Series Generative Adversarial Network) to preprocess, train, enhance data, and denoise the BCG signal through a deep learning model to improve the signal quality and temporal consistency of the data. The specific data processing flow is as follows:

[0122] Step 1: Use a BCG sensor (such as an intelligent mattress, pressure sensor) to obtain BCG signal data, including: ballistocardiogram signal waveform, timestamp, sensor noise;

[0123] The data is formatted into a time series data set, and each data record represents the BCG signal value at one time step.

[0124] To improve the signal quality and prevent the model from learning invalid noise information, the original data needs to be preprocessed, including:

[0125] Denoising: Use a band-pass filter (0.1Hz - 10Hz) to filter out power frequency interference and environmental noise.

[0126] Normalization: Adopt Z-score standardization processing to make the data mean 0 and variance 1 to improve the stability of model training.

[0127] Interpolation and Completion: For missing data points, linear interpolation or spline interpolation is used for completion to prevent data loss from affecting the training effect.

[0128] Time Alignment: Align the timing of the BCG signal to ensure that data from different subjects has the same time resolution (e.g., a sampling rate of 100 Hz).

[0129] Step 2: Select historical BCG data as the training set for the TimeGAN model, and the input format is temporally formatted BCG data. The random batch technique (mini-batch training) is adopted, and each batch contains 64 - 128 temporal samples to improve the training efficiency and the generalization ability of the model.

[0130] The TimeGAN model includes four core modules: Temporal Encoder, Generator, Temporal Decoder, and Discriminator;

[0131] Temporal Encoder: Use a two-layer GRU to extract the temporal features of the BCG signal and convert it into a latent space representation H real . The first layer of GRU extracts basic temporal features (heartbeat cycle, respiratory influence). The second layer of GRU captures longer-term temporal dependencies (such as heart rate variability). The fully connected layer (FC) maps to the latent space.

[0132] Generator: Generate the latent representation H using the random noise Z fake .

[0133] Structure: Two-layer GRU + self-attention mechanism to enhance the ability to model long-term dependencies.

[0134] Temporal Decoder: Restore H real and H fake back to the time-domain BCG signal. Among them, the data generated by Hfake can be used for data augmentation.

[0135] Discriminator: Responsible for distinguishing real temporal data from the generated temporal data to ensure the authenticity of the generated data.

[0136] Then use the Adam optimizer to optimize the Generator, Discriminator, Temporal Encoder, and Temporal Decoder respectively;

[0137] Step 3: Use the trained TimeGAN model to generate enhanced BCG data, and the steps are as follows:

[0138] Input the random noise Z, and generate the latent representation H through the Generator fake .

[0139] The timing decoder restores H fake to the enhanced BCG signal X fake .

[0140] Using the discriminator score, synthetic signals with higher quality are screened. The synthetic data is added to the original dataset to enhance data diversity and improve the model generalization ability.

[0141] Step 4: On the basis of the enhanced data, further denoise the BCG signal:

[0142] Perform a fast Fourier transform on X fake to transform it to the frequency domain and obtain the spectral component F fake .

[0143] Calculate the frequency attention weight W f to adaptively suppress the noise:

[0144]

[0145] where β is a preset steepness value; T is the threshold for setting noise suppression; |F fake | is the amplitude information of the BCG signal in the frequency domain.

[0146] Apply W f to F fake to suppress high-frequency noise and obtain the optimized F denoised .

[0147] Perform an inverse Fourier transform to convert F denoised back to the time domain and obtain the denoised enhanced BCG signal X denoised .

[0148] The adaptive frequency attention mechanism suppresses noise more accurately than traditional filtering methods and retains the effective signals. The data generated by TimeGAN conforms to the timing characteristics of real BCG signals, improving data diversity and authenticity. Using random batch training and Adam optimization speeds up the model training speed, making this method applicable to real-time medical monitoring systems. It can adapt to data from different acquisition devices and different populations, helping to improve the universality of the BCG signal analysis system.

[0149] In the description of the present invention, it should be understood that the terms "first" and "second" are only used for descriptive purposes and cannot be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of such features. In the description of the present invention, "a plurality of" means two or more unless otherwise specifically defined.

[0150] In the present invention, unless otherwise clearly specified or limited, the terms "installed", "connected", "connected to", "fixed", etc. shall be understood in a broad sense. For example, it may be a fixed connection, a detachable connection, or integrated; it may be a mechanical connection or an electrical connection; it may be directly connected or indirectly connected through an intermediate medium; it may be the communication inside two elements or the interaction relationship between two elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0151] In the present invention, unless otherwise clearly specified or limited, when the first feature is "on" or "under" the second feature, it may be that the first and second features are in direct contact, or the first and second features are indirectly in contact through an intermediate medium. Moreover, when the first feature is "above", "over" and "on top of" the second feature, it may be that the first feature is directly above or obliquely above the second feature, or simply means that the first feature is at a higher level than the second feature in terms of horizontal height. When the first feature is "under", "beneath" and "underneath" the second feature, it may be that the first feature is directly below or obliquely below the second feature, or simply means that the first feature is at a lower level than the second feature in terms of horizontal height.

[0152] In the description of this specification, the description of terms such as "one embodiment", "some embodiments", "embodiment", "example", "specific example" or "some examples" means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic descriptions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in a suitable manner in any one or more embodiments or examples. In addition, without contradiction, those skilled in the art can combine and combine the different embodiments or examples described in this specification and the features of different embodiments or examples.

[0153] Although the embodiments of the present invention have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those of ordinary skill in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for processing BCG signal data based on the TimeGAN model, characterized in that, Including: S1. Obtain the original BCG signal dataset X, and perform preprocessing on it to obtain time-series formatted BCG data; The preprocessing includes denoising, interpolation and completion, normalization, and time alignment; S2. Input the time-order formatted BCG data into the trained TimeGAN model to generate the enhanced BCG signal X fake ; S3. Perform noise suppression processing on the enhanced BCG signal X fake to obtain the denoised enhanced BCG signal X denoised .

2. According to the BCG signal data processing method based on the TimeGAN model described in claim 1, the S3 specifically includes: S31. Use the fast Fourier transform to transform the enhanced BCG signal X fake to the frequency domain to obtain the spectral component F fake ; S32. Calculate the frequency attention weight W f and apply it to the spectral component F fake to suppress high-frequency noise and obtain the optimized frequency-domain signal F denoised ; S33. Use the inverse Fourier transform to convert the frequency-domain signal F denoised back to the time domain to obtain the enhanced BCG signal X after denoising denoised .

3. According to the BCG signal data processing method based on the TimeGAN model described in claim 1, before the S1, it also includes: S0. Use the training dataset to train the pre-obtained TimeGAN model to obtain a trained TimeGAN model; The training dataset includes: time-series formatted BCG data for training the TimeGAN model in the historical time period; The TimeGAN model includes: A temporal encoder, which is used to receive the temporally formatted BCG data input into the TimeGAN model, extract temporal features, and convert them into a latent space representation H real ; A generator for generating a latent representation H from random noise Z fake ; A temporal decoder for converting a latent space representation H real and a latent representation H fake back into a temporal BCG signal, where the data restored from the latent representation H fake is the generated BCG data; A discriminator for distinguishing the latent space representation H real and the latent representation H fake and the difference between the time-series formatted BCG data and the generated BCG data.

4. According to the BCG signal data processing method based on the TimeGAN model described in claim 3, The time-series encoder uses two layers of gated recurrent units and one layer of fully connected layer for feature extraction. The first layer of gated recurrent unit extracts basic time-series characteristics, the second layer of gated recurrent unit captures complex time-series dependencies, and the fully connected layer is used to convert time-series features into latent space representations.

5. According to the BCG signal data processing method based on the TimeGAN model described in claim 2, Among them, Calculate the frequency attention weight W using formula (1). f ; The formula (1) is: where β is a preset steepness value; T is the threshold for setting noise suppression; |F fake |is the amplitude information of the BCG signal in the frequency domain.

6. According to the BCG signal data processing method based on the TimeGAN model described in claim 3, the TimeGAN model uses a self-supervised learning loss function during training; The self-supervised learning loss function is: L TimeGAN = γ1L rec + γ2L gen + γ3L adv + γ4L sup ; L rec is the reconstruction loss; L gen is the generation loss; L adv is the adversarial loss; L sup is the temporal consistency loss; γ1 is the first hyperparameter; γ2 is the second hyperparameter; γ3 is the third hyperparameter; γ4 is the fourth hyperparameter.

7. According to the BCG signal data processing method based on the TimeGAN model described in claim 6, wherein, Among them, is the reconstructed data generated by the encoder-decoder; E[] represents the expected sum of all data in []; L gen = E[log D gen (G(Z))]; Among them, G(Z) is the data generated by the generator; D gen is the discriminator's judgment on the synthetic data; L adv = E[log D(X) + log(1 - D(G(Z)))]; where D(X) is the discriminator's score for real data; D(G(Z)) is the score for generated data; where h t is the hidden state at time step t; is the hidden state obtained by the sequential encoder - sequential decoder.

8. According to the BCG signal data processing method based on the TimeGAN model described in claim 3, The TimeGAN model uses a random batch technique during training, and each batch of data includes 64 - 128 time-series formatted BCG data.

9. According to the BCG signal data processing method based on the TimeGAN model described in claim 8, During the training process of the TimeGAN model, a weighted adversarial loss function is used to enhance the temporal correlation between the generated BCG data and the time-series formatted BCG data; The weighted adversarial loss function is: L adv = E[log D(X real ))] + δ·E[log(1 - D(X fake ))]; L adv is the weighted adversarial loss function; D(·) is the output probability of the discriminator; X real is the true BCG signal; X fake is the generated BCG signal; δ is a preset dynamic weighting factor; E[] represents the expected sum of all data in [].

10. According to the BCG signal data processing method based on the TimeGAN model described in claim 9, During the training process of the TimeGAN model, the Adam optimizer is used to optimize the parameters of the generator, discriminator, time-series encoder, and time-series decoder to improve the quality and temporal consistency of the generated signal.

Images