BCG signal processing method and system based on LSTM and GAN fusion

Abstract

The invention relates to a BCG signal processing method and system based on LSTM and GAN fusion, and the method comprises the steps: obtaining a BCG original signal, and carrying out the signal preprocessing, and obtaining a preprocessed BCG signal; a time sequence feature coding module performs feature extraction on the preprocessed BCG signal to obtain time sequence feature data; based on confrontation training of an LSTM generator and a GAN discriminator, generating an enhanced BCG signal by using the LSTM generator according to the time sequence characteristic data, inputting the enhanced BCG signal into the GAN discriminator for authenticity evaluation, and if the GAN discriminator judges that the signal quality does not reach a preset standard, adjusting parameters of the LSTM generator and continuing training until the enhanced BCG signal meeting the quality requirement is generated; and inputting the enhanced BCG signal meeting the quality requirement into a signal noise reduction module to obtain a noise-reduced BCG signal, and performing signal reconstruction on the noise-reduced BCG signal by using a time sequence decoder to obtain a reconstructed BCG signal.

Description

Technical Field

[0001] The present invention relates to the technical field of signal processing, and in particular to a BCG signal processing method and system based on the fusion of LSTM and GAN. Background Art

[0002] BCG (Ballistocardiography) is a non-invasive cardiovascular function monitoring technology that records minute body movement signals caused by cardiac pulsation and blood flow. BCG signals are usually collected by high-precision pressure sensors, accelerometers or force sensors, and reflect the minute body movement changes caused by cardiac contraction and blood ejection.

[0003] However, BCG signals face many challenges in practical applications. First, due to the small amplitude of BCG signals, they are easily interfered by factors such as environmental noise, motion artifacts, and baseline drift, resulting in a decline in signal quality. Second, traditional signal processing methods, such as low-pass filtering and band-pass filtering, may cause loss of effective signals during the denoising process. In addition, due to factors such as individual differences and changes in sensor positions, the morphology of BCG signals may vary greatly, affecting the accuracy of subsequent signal analysis. Therefore, improving the quality of BCG signals and enhancing their stability is a key issue in current research. Summary of the Invention

[0004] In view of the above-mentioned disadvantages and deficiencies of the prior art, the present invention provides a BCG signal processing method and system based on the fusion of LSTM and GAN.

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

[0006] In a first aspect, an embodiment of the present invention provides a BCG signal processing method based on the fusion of LSTM and GAN, the method comprising:

[0007] S1. Obtain the BCG original signal, perform signal preprocessing to obtain the preprocessed BCG signal, and input the preprocessed BCG signal into the time series feature encoding module;

[0008] S2. The time series feature encoding module extracts features from the input preprocessed BCG signal to obtain time series feature data, and input the time series feature data into the LSTM generator;

[0009] S3. Based on the adversarial training of the LSTM generator and the GAN discriminator, use the LSTM generator to generate an enhanced BCG signal according to the time series feature data, and input the enhanced BCG signal into the GAN discriminator for authenticity evaluation. If the GAN discriminator determines that the signal quality does not meet the preset standard, adjust the parameters of the LSTM generator and continue training until an enhanced BCG signal that meets the quality requirements is generated;

[0010] S4. Input the enhanced BCG signal that meets the quality requirements into the signal denoising module, remove the residual noise through the adaptive noise suppression technology to obtain the denoised BCG signal, and use the time series decoder to reconstruct the denoised BCG signal to obtain the reconstructed BCG signal.

[0011] Preferably, the S1 specifically includes:

[0012] S11. Collect the BCG raw signal through a high-precision pressure sensor, and perform band-pass filtering on the BCG raw signal to remove baseline drift and high-frequency noise to obtain an initial signal;

[0013] S12. Use the wavelet transform method to decompose the initial signal, and perform threshold denoising processing using formula (1) to obtain the denoised BCG signal;

[0014] The formula (1) is:

[0015]

[0016] where T is the threshold; σ is the noise standard deviation; N is the number of signal sample points;

[0017] S13. Normalize the denoised BCG signal to map the signal amplitude to the range of [-1, 1] to obtain the normalized signal;

[0018] S14. Segment the normalized signal, with each segment length set to 256 sampling points, and use the sliding window method for segmentation, where the sliding window step size is 128 sampling points and the corresponding overlap rate is 50%, and finally obtain the preprocessed BCG signal.

[0019] Preferably, the S2 specifically includes:

[0020] S21. The time series feature encoding module uses a three-layer convolutional neural network to extract features from the preprocessed BCG signal to obtain a feature sequence;

[0021] Among them, the number of channels of each layer of the convolutional neural network is 64, 128, and 256 respectively, the convolutional kernel size is 3×1, and the step size is 1;

[0022] After each convolutional layer, batch normalization and ReLU activation functions are set to stabilize the training process and enhance the feature representation ability;

[0023] S22. Input the extracted feature sequence into the attention mechanism module to obtain enhanced temporal feature data;

[0024] The attention mechanism module uses 8 attention heads, and each attention head is used to calculate the weighted feature representation at different time steps;

[0025] S23. Input the enhanced temporal feature data into the LSTM generator.

[0026] Preferably, the S3 specifically includes:

[0027] S31. Use the LSTM generator to perform time step prediction on the input enhanced temporal feature data input into the LSTM generator to generate an enhanced BCG signal sequence;

[0028] The LSTM generator adopts a two-layer bidirectional LSTM structure, and the dimension of each hidden unit is set to 128;

[0029] S32. Use the generative adversarial network GAN to optimize the signal quality of the enhanced BCG signal sequence. If the GAN discriminator determines that the generated signal does not meet the quality standard, adjust the network parameters of the LSTM generator and continue the adversarial training until an enhanced BCG signal that meets the quality requirements is generated;

[0030] Among them, the training process of the generative adversarial network GAN adopts the min-max optimization strategy, and the loss function of the discriminator of the generative adversarial network GAN is:

[0031]

[0032] Among them, represents the score of the discriminator for the real data;

[0033] represents the score of the discriminator for the generated data;

[0034] D(x) represents the output of the discriminator for the real signal x;

[0035] represents the output of the discriminator for the generated signal ;

[0036] represents the expected operation of the real signal x sampled from the real data distribution P real ;

[0037] represents the data distribution P generated by the generator Ggen Sampled forged signal Expected operation;

[0038] logD(x) represents the logarithm of the probability that the discriminator correctly identifies the true signal x;

[0039] Represents the logarithm of the probability that the discriminator correctly identifies the forged signal ;

[0040] The quality standard includes: signal-to-noise ratio higher than 20dB, mean square error lower than 0.001, and dynamic time warping distance lower than 0.05.

[0041] Preferably, the S4 specifically includes:

[0042] S41. The signal denoising module uses a variational autoencoder VAE as a denoising autoencoder to remove residual noise in the enhanced BCG signal that meets the quality requirements through latent variable distribution learning, obtaining a denoised signal;

[0043] S42. Input the denoised signal into the temporal decoder to obtain a decoded signal;

[0044] The temporal decoder consists of two layers of bidirectional LSTM, and the hidden unit dimension is set to 128;

[0045] S43. Use transposed convolution to upsample the decoded signal to restore the decoded signal to the temporal structure of the original BCG signal, obtaining a reconstructed BCG signal.

[0046] Preferably, the method further includes:

[0047] S5. Perform quality assessment on the reconstructed BCG signal to obtain a quality assessment result.

[0048] Preferably,

[0049] The quality assessment result includes the signal-to-noise ratio value and mean square error value corresponding to the reconstructed BCG signal.

[0050] Preferably,

[0051] The optimization algorithm of the LSTM generator and the GAN discriminator uses the Adam optimizer;

[0052] The hyperparameters of the Adam optimizer are:

[0053] Learning rate: 0.0002;

[0054] First-order momentum parameter: 0.5;

[0055] Second-order momentum parameter: 0.999;

[0056] Batch size: 32;

[0057] Number of training epochs: 100.

[0058] Preferably,

[0059] S11 further includes performing adaptive filtering on the initial signal, where the cut-off frequency ωc of the band-pass filter is determined by formula (2);

[0060] The formula (2) is:

[0061]

[0062] S(f) is the power spectral density of the initial signal;

[0063] ω1 is the empirically set low-frequency cut-off value;

[0064] ω2 is the empirically set high-frequency cut-off value;

[0065] denotes finding the optimal cut-off frequency that maximizes the energy;

[0066] ω c is the cut-off frequency of the band-pass filter;

[0067] denotes the integral of the power spectral density S(f) of the initial signal over the frequency range [ω1, ω2].

[0068] On the other hand, this embodiment also provides a BCG signal processing system based on the fusion of LSTM and GAN, including:

[0069] At least one processor; and

[0070] At least one memory communicatively connected to the processor, wherein the memory stores program instructions executable by the processor, and the processor can execute the BCG signal processing method based on the fusion of LSTM and GAN as described in the first aspect by invoking the program instructions.

[0071] The beneficial effects of the present invention are:

[0072] A BCG signal processing method and system based on the fusion of LSTM and GAN according to the present invention can effectively improve the quality and integrity of the timing characteristics of the BCG signal due to the adoption of the adversarial training mechanism combining the LSTM generator and the GAN discriminator. Compared with the prior art, it can enhance the BCG signal under low signal-to-noise ratio conditions, improve the availability of the signal, and reduce the distortion problems caused by noise and signal loss. At the same time, through the joint feature extraction of the attention mechanism and the convolutional neural network, the ability to represent timing characteristics is enhanced, making the generated BCG signal more physiologically reasonable. It achieves the technical effects of improving the quality of the BCG signal, reducing noise interference, enhancing the signal timing consistency, and ultimately improving the analysis accuracy and reliability of the ballistocardiogram signal.

[0073] A BCG signal processing method and system based on the fusion of LSTM and GAN according to the present invention can effectively remove baseline drift and high-frequency noise through the signal preprocessing method combining wavelet transform threshold denoising method and band-pass filtering, making the BCG signal more stable and providing a better data basis for subsequent feature extraction and signal enhancement. Compared with the prior art, it can improve the feature extraction ability of the BCG signal through the timing feature encoding method combining a three-layer convolutional neural network and the attention mechanism, fully excavate the timing dependence relationship in the signal, and make the signal enhancement more physiologically credible. It achieves the technical effects of enhancing the ability to express the timing characteristics of the BCG signal, reducing the loss of key physiological information, and improving the accuracy after signal reconstruction, providing more reliable data support for the intelligent analysis and application of the ballistocardiogram signal.

[0074] In addition, due to the use of a bidirectional LSTM generator for signal prediction and combining GAN adversarial training to optimize the signal quality, it can adaptively generate high-quality BCG signals, effectively suppress false signals, and improve the stability of signal enhancement. Compared with the prior art, it can effectively improve the quality of the BCG signal under small sample conditions, reduce the influence of the measurement environment on signal acquisition, and improve the availability of low-quality signals, thereby enhancing the adaptability and wide applicability of the BCG signal. It achieves the technical effects of improving the quality and robustness of the BCG signal under different measurement conditions, enhancing the practical value of the signal, and making the clinical analysis and health monitoring of the ballistocardiogram signal more accurate and reliable. Description of the Drawings

[0075] Figure 1 It is a flowchart of a BCG signal processing method based on the fusion of LSTM and GAN according to the present invention;

[0076] Figure 2 It is a flowchart of a BCG signal processing method based on the fusion of LSTM and GAN in an embodiment of the present invention. Detailed Embodiments

[0077] For better explaining the present invention for easier understanding, the present invention will be described in detail below in conjunction with the accompanying drawings through specific embodiments.

[0078] To better understand the above technical solution, 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 more clear and thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.

[0079] Embodiment 1

[0080] See Figure 1 , this embodiment provides a BCG signal processing method based on the fusion of LSTM and GAN, and the method includes:

[0081] S1. Obtain the BCG original signal, perform signal preprocessing to obtain the preprocessed BCG signal, and input the preprocessed BCG signal into the time series feature encoding module;

[0082] In this embodiment, it is assumed that a certain medical institution uses a high-precision pressure sensor to collect the BCG signal of a patient under the hospital mattress. Due to factors such as patient body movement and mattress elasticity, the collected original signal contains baseline drift and high-frequency noise. To improve the signal quality, band-pass filtering (e.g., 0.1Hz - 10Hz) is used to remove baseline drift and high-frequency interference. Subsequently, wavelet transform is used to decompose the signal, and the threshold denoising method is used to remove random noise. After denoising, normalization processing is performed to map the signal amplitude to the range of [-1, 1] to reduce data deviation caused by individual differences. Finally, the signal is segmented by a sliding window (window length 256 sampling points, step size 128 sampling points) to form continuous time series data for subsequent feature extraction.

[0083] In the practical application of this embodiment, the S1 specifically includes:

[0084] S11. Collect the BCG original signal through a high-precision pressure sensor, and perform band-pass filtering on the BCG original signal to remove baseline drift and high-frequency noise to obtain an initial signal;

[0085] S12. Decompose the initial signal by using the method of wavelet transform, and perform threshold denoising processing using formula (1) to obtain the denoised BCG signal;

[0086] The formula (1) is:

[0087]

[0088] Among them, T is the threshold; σ is the standard deviation of the noise; N is the number of signal sample points;

[0089] S13. Normalize the denoised BCG signal to map the signal amplitude to the range of [-1, 1] to obtain the normalized signal;

[0090] S14. Segment the normalized signal, with each segment having a length of 256 sampling points, and use a sliding window method for segmentation, where the sliding window step size is 128 sampling points and the corresponding overlap rate is 50%, and finally obtain the preprocessed BCG signal.

[0091] S2. The temporal feature encoding module extracts features from the input preprocessed BCG signal to obtain temporal feature data, and inputs the temporal feature data into the LSTM generator;

[0092] Specifically, the S2 specifically includes:

[0093] S21. The temporal feature encoding module uses a three-layer convolutional neural network to extract features from the preprocessed BCG signal to obtain a feature sequence;

[0094] Among them, the number of channels of each layer of the convolutional neural network is 64, 128, and 256 respectively, the convolutional kernel size is 3×1, and the step size is 1;

[0095] After each convolutional layer, batch normalization and ReLU activation functions are set to stabilize the training process and enhance the feature representation ability;

[0096] S22. Input the extracted feature sequence into the attention mechanism module to obtain enhanced temporal feature data;

[0097] The attention mechanism module uses 8 attention heads, and each attention head is used to calculate the weighted feature representation at different time steps;

[0098] S23. Input the enhanced temporal feature data into the LSTM generator.

[0099] For example, a three-layer convolutional neural network (CNN) is used to extract temporal features from the preprocessed signal. The specific process is as follows: The first convolutional layer: Set 64 channels, the convolutional kernel size is 3×1, the stride is 1, and low-level features (such as local fluctuations) are extracted. The second convolutional layer: Set 128 channels to extract deeper features (such as rhythm patterns). The third convolutional layer: Set 256 channels to further capture long-term correlation features. Batch normalization + ReLU activation function: Ensure training stability and enhance non-linear features. Set 8 attention heads to make the network focus on the key parts of the BCG signal (such as the impact signal during heart pumping). In this embodiment, key features are automatically extracted without manual screening, improving the automation level of BCG signal analysis. The attention mechanism enhances the expression of key physiological signals and improves the model's attention to important information.

[0100] S3. Based on the adversarial training of the LSTM generator and the GAN discriminator, use the LSTM generator to generate an enhanced BCG signal according to the temporal feature data, and input the enhanced BCG signal into the GAN discriminator for authenticity evaluation. If the GAN discriminator determines that the signal quality does not meet the preset standard, adjust the parameters of the LSTM generator and continue training until an enhanced BCG signal that meets the quality requirements is generated;

[0101] Specifically, the S3 specifically includes:

[0102] S31. Input the enhanced temporal feature data into the LSTM generator through the LSTM generator for time step prediction to generate an enhanced BCG signal sequence;

[0103] The LSTM generator adopts a two-layer bidirectional LSTM structure, and the dimension of each hidden unit is set to 128;

[0104] S32. Use the generative adversarial network GAN to optimize the signal quality of the enhanced BCG signal sequence. If the GAN discriminator determines that the generated signal does not meet the quality standard, adjust the network parameters of the LSTM generator and continue the adversarial training until an enhanced BCG signal that meets the quality requirements is generated;

[0105] Among them, the training process of the generative adversarial network GAN adopts the min-max optimization strategy, and the loss function of the discriminator of the generative adversarial network GAN is:

[0106]

[0107] Among them, represents the discriminator's score for real data; represents the discriminator's score for generated data; D(x) represents the discriminator's output for the real signal x; represents the discriminator's score for the generated signal Output; Denotes the expected operation on the true signal x sampled from the true data distribution P real Sampled true signal x; Denotes the expected operation on the forged signal sampled from the data distribution P generated by the generator G gen Sampled forged signal Expected operation; logD(x) denotes the logarithm of the probability that the discriminator correctly identifies the true signal x; Denotes the logarithm of the probability that the discriminator correctly identifies the forged signal Probability value;

[0108] The quality criteria include: signal-to-noise ratio higher than 20 dB, mean square error lower than 0.001, and dynamic time warping distance lower than 0.05.

[0109] For example, the LSTM generator: adopts a bidirectional LSTM structure to predict the BCG signal at the next time step and generate an enhanced signal. The GAN discriminator: is used to determine whether the enhanced signal is consistent with the true signal. Training process: If the GAN discriminator determines that the quality of the generated signal is poor (such as signal-to-noise ratio < 20 dB, mean square error > 0.001), then adjust the parameters of the LSTM generator to make it generate a more realistic signal. Train until the quality of the generated BCG signal meets the medical application standards (such as the dynamic time warping distance close to the true data is lower than 0.05).

[0110] In this embodiment, the adversarial training ensures continuous improvement of the signal quality and avoids generating false information during the signal enhancement process. The LSTM generator utilizes the long-term and short-term memory characteristics to ensure the physiological rationality of the signal enhancement and make it closer to the true BCG signal.

[0111] S4. Input the enhanced BCG signal that meets the quality requirements into the signal denoising module, remove the residual noise through the adaptive noise suppression technology to obtain the denoised BCG signal, and use the temporal decoder to reconstruct the denoised BCG signal to obtain the reconstructed BCG signal.

[0112] Specifically, S4 specifically includes:

[0113] S41. The signal denoising module uses the variational autoencoder VAE as a denoising autoencoder to remove the residual noise in the enhanced BCG signal that meets the quality requirements through the latent variable distribution learning to obtain the denoised signal;

[0114] S42. Input the denoised signal into the temporal decoder to obtain the decoded signal;

[0115] The temporal decoder consists of two layers of bidirectional LSTM, and the hidden unit dimension is set to 128;

[0116] S43. Upsample the decoded signal in a deconvolution manner to restore the decoded signal to the timing structure of the original BCG signal, obtaining the reconstructed BCG signal.

[0117] In this embodiment, a BCG signal processing method and system based on the fusion of LSTM and GAN can effectively remove baseline drift and high-frequency noise by adopting the joint denoising strategy of wavelet transform and band-pass filtering, making the signal clearer and more stable. Compared with the prior art, it can improve the automation level of signal processing and enhance the ability to express the timing characteristics of the signal through the timing feature extraction method of deep learning, making it closer to the real BCG signal. Through the adversarial training mechanism of LSTM and GAN, this method can intelligently generate high-quality BCG signals, improve the usability of low-quality signals, reduce the influence of the measurement environment on signal acquisition, and thus enhance the practical application value of BCG signals. Finally, through the adaptive noise suppression and signal reconstruction methods, the processed BCG signal is more stable and reliable, achieving the technical effects of improving the quality of BCG signals, enhancing the accuracy of ballistocardiogram signal analysis, and making it more valuable in applications such as health monitoring and clinical diagnosis.

[0118] See Figure 2 , in this embodiment, the method further includes:

[0119] S5. Perform quality assessment on the reconstructed BCG signal to obtain a quality assessment result.

[0120] The quality assessment result includes the signal-to-noise ratio value and the mean square error value corresponding to the reconstructed BCG signal.

[0121] In this embodiment, the optimization algorithm of the LSTM generator and the GAN discriminator adopts the Adam optimizer; the hyperparameters of the Adam optimizer are:

[0122] Learning rate: 0.0002; first-order momentum parameter: 0.5; second-order momentum parameter: 0.999; batch size: 32; number of training epochs: 100.

[0123] Adam is an adaptive gradient optimization algorithm that combines the advantages of Momentum optimization and RMSProp. It can efficiently adjust the learning rate, making the model converge more stably and the training more efficient. In this embodiment, a relatively small learning rate of 0.0002 is selected, which can prevent gradient explosion or gradient disappearance during GAN training and improve the balance between the generator and the discriminator. It is applicable to time series data learning (such as BCG signals), avoiding the decline in the quality of the generated signal caused by too fast model updates. Lower than the default value of 0.9, it reduces the influence of old gradients, improves the stability of GAN training, and prevents mode collapse. It makes the generator and the discriminator more balanced during training, avoiding unstable oscillations of GAN at the initial stage of training. A relatively high second-order momentum parameter is adopted to ensure the stability of the estimated value of the gradient square and prevent excessive adjustment of the learning rate during training. Mini-batch training can improve the generalization ability of the model, accelerate the calculation at the same time, and improve GPU utilization. It is suitable for BCG signal data training because too large a batch may cause the model to be difficult to capture small physiological changes. 100 rounds of training can ensure that the model fully learns the data features and generates high-quality enhanced BCG signals. If the number of rounds is too small, the model may be underfitted, resulting in a lower quality of the generated signal; if too many, overfitting problems may occur.

[0124] In this embodiment, the S11 further includes performing adaptive filtering processing on the initial signal, where the cut-off frequency ω of the band-pass filter c is determined by using formula (2);

[0125] The formula (2) is:

[0126]

[0127] S(f) is the power spectral density of the initial signal;

[0128] ω1 is the empirically set low-frequency cut-off value;

[0129] ω2 is the empirically set high-frequency cut-off value;

[0130] represents finding the optimal cut-off frequency that maximizes the energy;

[0131] ω c is the cut-off frequency of the band-pass filter;

[0132] represents the integral of the power spectral density S(f) of the initial signal in the frequency range [ω1, ω2].

[0133] For example, the initial signal usually contains physiological noise (such as respiratory artifacts) and environmental noise (such as electromagnetic interference). To enhance the BCG signal, a band-pass filter is used, but how to select the appropriate cut-off frequency ωc is crucial.

[0134] In this embodiment, an empirical range is set. For example, the low-frequency cut-off value ω1 = 0.5 Hz (to filter out ultra-low-frequency noise such as the influence of breathing). The high-frequency cut-off value ω2 = 15.0 Hz (to avoid electromagnetic interference at too high frequencies). Formula (2) is used for optimization. At this time, formula (2) means finding an optimal band-pass cut-off frequency within the range of [0.5 Hz, 15.0 Hz] to maximize the signal energy. Through calculation, it is obtained that between 2 Hz and 10 Hz, the proportion of the BCG signal energy is the largest. Therefore, the cut-off frequency of the band-pass filter can be optimized to ω c = [2, 10] Hz, which can maximize the signal energy while reducing the loss of useful information.

[0135] This embodiment also provides a BCG signal processing system based on the fusion of LSTM and GAN, including: at least one processor; and at least one memory communicatively connected to the processor, wherein the memory stores program instructions executable by the processor, and the processor can execute the BCG signal processing method based on the fusion of LSTM and GAN as described in Embodiment 1 by calling the program instructions.

[0136] Embodiment 2

[0137] This Embodiment 2 provides a BCG signal processing method based on the fusion of LSTM and GAN, specifically including:

[0138] Step 101: First, collect the BCG signal using a high-precision pressure sensor or a fiber optic sensor. The sampling frequency is set to 1000 Hz to ensure sufficient time resolution. The collected original signal is stored as a two-dimensional matrix: X raw = [x1, x2,..., x n ; x i ∈R;

[0139] where X raw is the original BCG signal, and x i is the signal value at the i-th sampling point.

[0140] Step 102: Use a Butterworth low-pass filter (cut-off frequency 30 Hz) to remove high-frequency noise; use a high-pass filter (cut-off frequency 0.5 Hz) to remove low-frequency drift; use wavelet transform (Daubechies-4 wavelet) for multi-scale denoising;

[0141] Step 103: Since the BCG signals of different users may drift on the time axis, the signals need to be aligned. The electrocardiogram R-wave alignment method is adopted to detect the RR interval of the heartbeat cycle according to the ECG signal, and then the BCG signal is re-cut so that each segment of the signal corresponds to a complete cardiac cycle.

[0142] Step 104: To avoid the influence of differences in signal amplitudes on training, the signals are normalized to [-1, 1];

[0143] The processed signals are divided according to a sliding window (window length 256, overlap 50%) to form data batches and stored in the tensor X batch as the input for subsequent LSTM-GAN training.

[0144] Step 106: The BCG signal processed in Step 105 enters the temporal encoder to extract temporal features and generate a feature vector H enc ;

[0145] H enc = f enc (X batch );

[0146] where f enc () consists of 3 layers of CNN + batch normalization + max pooling, and the number of channels in each layer is 64, 128, and 256 respectively.

[0147] Step 107: The feature vector H enc is generated into a new signal H gen through the LSTM generator;

[0148] H gen = f gen (H enc );

[0149] where f gen () adopts 2 layers of bidirectional LSTM (128-dimensional hidden units) + multi-head attention mechanism (8 heads) + residual connection.

[0150] Step 108: The discriminator discriminates the authenticity of the generated signal:

[0151] P real / fake = f dis (H gen )

[0152] where f dis () uses 3 layers of CNN + 1 layer of LSTM for feature extraction and temporal analysis, and outputs the true / false discrimination probability P real / fake .

[0153] Step 109: The timing decoder performs signal reconstruction to obtain the reconstructed BCG signal.

[0154] Step 110: Calculate the perceptual evaluation of speech quality (PESQ) of the final signal to ensure that the quality meets the standard. Calculate the signal statistical distribution to ensure that the generated signal is consistent with the real signal distribution. Adopt the parameter adaptive adjustment technology to dynamically adjust the LSTM-GAN training parameters according to the quality evaluation results.

[0155] In this embodiment, there is strong environmental noise (such as mechanical vibration, mattress deformation, etc.) in the original BCG signal, which directly affects the signal quality. This solution uses Butterworth filtering, wavelet denoising, and LMS adaptive filtering, and combines LSTM-GAN to generate enhanced signals, thereby significantly improving the signal-to-noise ratio (SNR).

[0156] This embodiment adopts data augmentation + adaptive GAN training, which can be adapted to different types of BCG sensors (such as MEMS accelerometers, optical BCG, capacitive BCG).

[0157] 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 understood as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, the 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" means two or more, unless otherwise specifically defined.

[0158] In the present invention, unless otherwise clearly defined and limited, the terms "installed", "connected", "connected", "fixed", etc. should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or integrated; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium; it can be the internal connection of two components or the interaction relationship between two components. 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 situations.

[0159] In the present invention, unless otherwise clearly defined and limited, the first feature being "on" or "under" the second feature can 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, the first feature being "above", "over" and "on" the second feature can be that the first feature is directly above or obliquely above the second feature, or simply means that the first feature has a higher horizontal height than the second feature. The first feature being "under", "below" and "beneath" the second feature can be that the first feature is directly below or obliquely below the second feature, or simply means that the first feature has a lower horizontal height than the second feature.

[0160] In the description of this specification, the descriptions of terms such as "one embodiment", "some embodiments", "embodiment", "example", "specific example", or "some examples" mean 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 representations 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.

[0161] 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 BCG signal processing method based on the fusion of LSTM and GAN, characterized in that The method includes: S1. Obtain the BCG original signal, perform signal preprocessing to obtain the preprocessed BCG signal, and input the preprocessed BCG signal into the temporal feature encoding module; S2. The temporal feature encoding module extracts features from the input preprocessed BCG signal to obtain temporal feature data, and input the temporal feature data into the LSTM generator; S3. Based on the adversarial training of the LSTM generator and the GAN discriminator, use the LSTM generator to generate an enhanced BCG signal according to the temporal feature data, and input the enhanced BCG signal into the GAN discriminator for authenticity evaluation. If the GAN discriminator determines that the signal quality does not meet the preset standard, adjust the parameters of the LSTM generator and continue training until an enhanced BCG signal that meets the quality requirements is generated; S4. Input the enhanced BCG signal that meets the quality requirements into the signal denoising module, remove the residual noise through the adaptive noise suppression technology to obtain the denoised BCG signal, and use the temporal decoder to reconstruct the denoised BCG signal to obtain the reconstructed BCG signal.

2. The BCG signal processing method based on the fusion of LSTM and GAN according to claim 1, wherein The specific content of S1 includes: S11. Collect the BCG original signal through a high-precision pressure sensor, and perform band-pass filtering on the BCG original signal to remove baseline drift and high-frequency noise to obtain an initial signal; S12. Decompose the initial signal by using the wavelet transform method, and perform threshold denoising processing using formula (1) to obtain the denoised BCG signal; The formula (1) is: where T is the threshold; σ is the noise standard deviation; N is the number of signal sample points; S13. Normalize the denoised BCG signal to map the signal amplitude to the range of [-1, 1] to obtain the normalized signal; S14. Segment the normalized signal, with each segment having a length of 256 sampling points, and perform segmentation in a sliding window manner, where the sliding window step size is 128 sampling points and the corresponding overlap rate is 50%, and finally obtain the preprocessed BCG signal.

3. The BCG signal processing method based on the fusion of LSTM and GAN according to claim 1, wherein, The specific content of S2 includes: S21. The temporal feature encoding module uses a three-layer convolutional neural network to extract features from the preprocessed BCG signal to obtain a feature sequence; Among them, the number of channels of each layer of the convolutional neural network is 64, 128, and 256 respectively, the convolutional kernel size is 3×1, and the step size is 1; A batch normalization and ReLU activation function are set after each convolutional layer to stabilize the training process and enhance the feature representation ability; S22. Input the extracted feature sequence into the attention mechanism module to obtain enhanced temporal feature data; The attention mechanism module uses 8 attention heads, and each attention head is used to calculate the weighted feature representation at different time steps; S23. Input the enhanced temporal feature data into the LSTM generator.

4. The BCG signal processing method based on the fusion of LSTM and GAN according to claim 1, characterized in that, The specific content of S3 includes: S31. Input the enhanced temporal feature data into the LSTM generator through the LSTM generator for time step prediction to generate an enhanced BCG signal sequence; The LSTM generator adopts a two-layer bidirectional LSTM structure, and the dimension of the hidden units in each layer is set to 128; S32. Use a generative adversarial network (GAN) to optimize the signal quality of the enhanced BCG signal sequence. If the GAN discriminator determines that the generated signal does not meet the quality standard, adjust the network parameters of the LSTM generator and continue the adversarial training until an enhanced BCG signal that meets the quality requirements is generated; Among them, the training process of the generative adversarial network (GAN) adopts the min-max optimization strategy, and the loss function of the discriminator of the generative adversarial network (GAN) is: Among them, represents the discriminator's score for real data; Indicates the discriminator's score for the generated data; D(x) represents the output of the discriminator for the real signal x; Indicates the output of the discriminator for the generated signal ; Denote the expected operation on the true signal x sampled from the true data distribution P real ; denote the data distribution P generated by the generator G gen sampled forged signal expectation operation; logD(x) represents the logarithmic value of the probability that the discriminator correctly identifies the real signal x; The logarithmic value of the probability that the discriminator correctly identifies the forged signal ; The quality standard includes: the signal-to-noise ratio is higher than 20 dB, the mean square error is lower than 0.001, and the dynamic time warping distance is lower than 0.

05.

5. The BCG signal processing method based on the fusion of LSTM and GAN according to claim 1, wherein The specific steps of S4 are as follows: S41. The signal denoising module uses a variational autoencoder (VAE) as a denoising autoencoder to remove the residual noise in the enhanced BCG signal that meets the quality requirements through learning the latent variable distribution, and obtain the denoised signal; S42. Input the denoised signal into the temporal decoder to obtain the decoded signal; The temporal decoder consists of two layers of bidirectional LSTM, and the dimension of the hidden units is set to 128; S43. Use the transposed convolution method to upsample the decoded signal to restore the decoded signal to the temporal structure of the original BCG signal, and obtain the reconstructed BCG signal.

6. The BCG signal processing method based on the fusion of LSTM and GAN according to claim 1, wherein The method further includes: S5. Perform quality evaluation on the reconstructed BCG signal to obtain a quality evaluation result.

7. The BCG signal processing method based on the fusion of LSTM and GAN according to claim 6, characterized in that The quality evaluation result includes the signal-to-noise ratio value and the mean square error value corresponding to the reconstructed BCG signal.

8. The BCG signal processing method based on the fusion of LSTM and GAN according to claim 1, characterized in that The optimization algorithm of the LSTM generator and the GAN discriminator adopts the Adam optimizer; The hyperparameters of the Adam optimizer are: Learning rate: 0.0002; First-order momentum parameter: 0.5; Second-order momentum parameter: 0.999; Batch size: 32; Number of training epochs:

100.

9. The BCG signal processing method based on the fusion of LSTM and GAN according to claim 2, characterized in that The S11 further includes performing adaptive filtering processing on the initial signal, and the cut-off frequency ωc of the band-pass filter is determined by formula (2); The formula (2) is: S(f) is the power spectral density of the initial signal; ω1 is the empirically set low-frequency cut-off value; ω2 is the empirically set high-frequency cut-off value; Indicates finding the optimal cut-off frequency for maximizing energy; ω c is the cut-off frequency of the band-pass filter; Denotes the integral of the power spectral density S(f) of the initial signal over the frequency interval [ω1, ω2].

10. A BCG signal processing system based on the fusion of LSTM and GAN, characterized in that, including: At least one processor; and At least one memory communicatively connected to the processor, wherein the memory stores program instructions executable by the processor, and the processor can execute the BCG signal processing method based on the fusion of LSTM and GAN according to any one of claims 1-9 by invoking the program instructions.

Images