Intelligent heart sound classification method based on real-time streaming signal denoising and bispectrum characteristics

Abstract

The application discloses an intelligent heart sound classification method based on real-time streaming signal denoising and bispectrum characteristics, and comprises the following steps: S1, collecting heart sound signals through a heart sound sensor, amplifying and filtering the initial signals, and completing envelope extraction and digitization conversion of the heart sound signals; S2, pre-processing and intelligently denoising the digital signals, and extracting bispectrum characteristics of the denoised pure digital signals; S3, inputting the bispectrum characteristic two-dimensional image tensor into a trained heart sound intelligent classification network model, and outputting the probability distribution of five states of normal heart sound, aortic valve stenosis, mitral valve regurgitation, mitral valve stenosis and mitral valve prolapse. According to clinical data verification, the accuracy rate of the application in the abnormal heart sound recognition task reaches 96%, and the accuracy rate is improved by about 5% compared with similar algorithms.

Description

Technical Field

[0001] This invention belongs to the field of signal processing technology, specifically relating to an intelligent heart sound classification method based on real-time streaming signal denoising and bispectral features. Background Technology

[0002] Early screening for cardiovascular diseases relies primarily on the analysis of acoustic signals related to cardiac mechanical activity. While auscultation of heart sounds is the preferred non-invasive method for detecting valvular lesions and congenital heart disease, traditional auscultation techniques face the dual challenges of a lack of objective standards and severe noise interference in clinical applications. Although existing electronic auscultation devices have digitized the signals, they are mostly limited to time-domain waveform playback or linear frequency-domain analysis based on Fast Fourier Transform (FFT).

[0003] However, biomedical signals exhibit significant nonstationarity and nonlinearity. The physical mechanisms of heart sound generation involve valve opening and closing, blood flow turbulence, and ventricular wall vibration. These mechanical activities demonstrate complex nonlinear interactions, containing rich quadratic phase coupling (QPC) information. Traditional linear analysis methods such as power spectrum or Mel-Cepstrum analysis utilize only second-order statistics, inevitably losing the aforementioned crucial phase coupling information. This leads to a sharp decline in feature extraction capabilities in low signal-to-noise ratio environments, making it impossible to accurately distinguish pathological signals with similar spectral characteristics, such as aortic stenosis and mitral regurgitation. Furthermore, existing systems mostly employ an offline architecture of "acquisition-end storage-host computer analysis," lacking the ability to achieve end-to-end real-time high-precision diagnosis on resource-constrained embedded edge devices. Summary of the Invention

[0004] This invention addresses the shortcomings of existing technologies by providing an intelligent heart sound classification method based on real-time streaming signal denoising and bispectral features. Through high-order spectral analysis, it accurately identifies heart sound signals with similar features, overcoming the limitations of simple signal analysis methods and achieving accurate identification of abnormal heart sounds. Furthermore, it solves the technical challenge of real-time streaming heart sound signals being easily obscured by environmental noise and motion artifacts through efficient artificial intelligence algorithms.

[0005] This invention achieves this objective through the following technical solution: This invention provides an intelligent heart sound classification method based on real-time streaming signal denoising and bispectral features, comprising the following steps: S1: Collect heart sound signals through a heart sound sensor, amplify and filter the initial signal, and complete the envelope extraction and digital conversion of the heart sound signal; S2: Perform preprocessing and intelligent denoising on the digital signal, and extract bispectral features from the denoised clean digital signal; S3: Input the bispectral feature two-dimensional image tensor into the trained heart sound intelligent classification network model, and output the probability distribution of five states: normal heart sound, aortic stenosis, mitral regurgitation, mitral stenosis and mitral prolapse.

[0006] Furthermore, in step S1, the initial heart sound signal is first amplified and bandpass filtered, then amplified further and digitized by an MCU chip. An envelope extraction circuit is set between the primary and secondary amplification circuits to set the time constant. .

[0007] Furthermore, the envelope extraction circuit sets a time constant. It takes 103ms.

[0008] Furthermore, the preprocessing in step S2 includes the following: First, resampling is performed to uniformly resample the input signal to 1000Hz; Secondly, digital bandpass filtering is performed, and secondary filtering is carried out in the range of 25~400Hz. Then, normalization is performed to normalize the signal amplitude to the range of [-1, 1]. Finally, heart sound segmentation is performed, and the energy envelope method is used to locate the cardiac cycle. The first and second heart sounds are located by segmentation algorithm. Then, signal segments containing at least two complete heartbeat cycles are extracted for subsequent intelligent denoising and high-order spectral feature calculation and analysis.

[0009] Furthermore, the intelligent denoising process is as follows: the streaming heart sound signal is processed in frames, and each frame is input into the intelligent denoising U-Net network. The network estimates the background noise component in the frame and subtracts it from the original signal to obtain a preliminary cleaned signal. Subsequently, the preliminary cleaned signal and its corresponding noise estimate are transformed to the frequency domain, the noise power spectrum is updated by recursive smoothing, and the Wiener gain is calculated to perform amplitude correction on the spectrum. Finally, the continuous time domain signal is reconstructed by inverse Fourier transform and overlapping addition method to achieve low-latency, high-fidelity streaming denoising.

[0010] Furthermore, the intelligent denoising U-Net network model structure is as follows: The network adopts a one-dimensional U-Net architecture, and realizes end-to-end heart sound signal denoising through a symmetrical encoder-decoder structure. The encoder downsamples layer by layer to extract multi-scale features, and introduces a self-attention mechanism after each resolution to model global dependencies; the decoder gradually recovers the signal length through upsampling, and at the same time uses skip connections to fuse the fine-grained details of the encoder with the semantic features of the decoder, thereby effectively suppressing noise while preserving the fine temporal structure of the signal, and finally outputting a clean heart sound waveform.

[0011] Furthermore, during bispectral feature extraction, the preprocessed continuous time-domain heart sound signal is divided into frames, with a frame length of [missing information]. ,overlapping This yields a 256×256 two-dimensional bispectral matrix. The amplitude values ​​of the bispectral matrix are linearly normalized to [0, 255], and then reshaped into a two-dimensional image tensor of size 1×256×256, which is used as the input of the deep learning classification model.

[0012] Furthermore, the intelligent heart sound classification network model structure includes two cascaded convolutional blocks. The first convolutional block is configured with 32 3×3 convolutional kernels, and the second convolutional block is configured with 64 3×3 convolutional kernels. Each convolutional layer is followed by a ReLU activation function and a 2×2 max pooling layer to extract texture and shape features from the bispectral map. The extracted high-dimensional feature vector is input to a fully connected layer containing 128 neurons. Finally, the probability distribution for five states—normal heart sound, aortic stenosis, mitral regurgitation, mitral stenosis, and mitral prolapse—is output through a Softmax layer.

[0013] This invention also provides an intelligent heart sound classification device based on real-time streaming signal denoising and high-order spectral feature extraction, comprising: The heart sound sensing and acquisition module is used to acquire the initial heart sound signal and complete the digital conversion; The real-time heart sound signal intelligent denoising and feature extraction module uses a built-in algorithm to denoise the real-time heart sound digital signal and extract the bispectral features of the denoised heart sound digital signal; The intelligent classification and diagnosis module classifies the input bispectral features using a trained deep learning intelligent classification model and outputs the probability distribution of five states: normal heart sounds, aortic stenosis, mitral regurgitation, mitral stenosis, and mitral valve prolapse.

[0014] Furthermore, by connecting to an intelligent language model platform, it interacts with users and provides personalized suggestions based on their interactions.

[0015] Compared with the prior art, the beneficial effects of this invention are as follows: This technical solution effectively overcomes the deficiency of phase information loss in traditional linear spectral analysis by introducing higher-order spectral analysis, significantly improving the ability to capture weak pathological features in complex acoustic environments. A dual noise reduction mechanism ensures high fidelity of the features. Clinical data validation shows that this invention achieves an accuracy of 96% in abnormal heart sound recognition, representing an improvement of approximately 5% compared to similar algorithms. Attached Figure Description

[0016] Figure 1 This is a flowchart of the present invention; Figure 2This is a schematic diagram of the heart sound sensing and acquisition module of the present invention; Figure 3 This is a schematic diagram of the intelligent denoising U-Net network model structure of the present invention; Figure 4 This is a flowchart of the digital signal processing and bispectral feature extraction process of the present invention; Figure 5 This is a structural diagram of the intelligent heart sound classification network model of the present invention; Figure 6 This is a comparison chart of the noise reduction performance test results of the present invention. Detailed Implementation

[0017] Exemplary embodiments of the present invention will now be described in more detail with reference to the accompanying drawings. While 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 to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present invention and to fully convey the scope of the invention to those skilled in the art. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0018] See Figures 1 to 6 This invention discloses an intelligent heart sound classification method and device based on real-time streaming signal denoising and high-order spectral feature extraction. The device includes a heart sound sensing and acquisition module, a real-time intelligent heart sound signal denoising and feature extraction module, and an intelligent classification and diagnosis module. Through these three modules, the system can acquire heart sound signals in real time and perform intelligent classification of the heart sound signals in real time. While providing preliminary diagnosis, the system can also interact with users through an intelligent language model platform to provide more personalized suggestions.

[0019] The process of this invention is as follows Figure 1 As shown, the specific steps are as follows: S1: Collect heart sound signals through a heart sound sensor, amplify and filter the initial signal, and complete the envelope extraction and digital conversion of the heart sound signal.

[0020] This step is implemented through the heart sound sensing and acquisition module.

[0021] As attached Figure 2 As shown, the heart sound sensing and acquisition module includes a heart sound sensor and a signal conditioning circuit. The heart sound sensor employs a high-sensitivity, low-noise piezoelectric sensor, with a frequency response range covering the core heart sound frequency band (20~150Hz). Optimized acoustic coupling design ensures that the acoustic signal is converted into a high-quality analog electrical signal to the greatest extent possible. Environmental noise interference is reduced through mechanical isolation and pre-shielding.

[0022] Signal input: Heart sound sensor interface (3.5mm headphone jack), HKY-06B+ analog heart sound sensor, input via headphone jack, main frequency components are concentrated in 20~150Hz.

[0023] The signal conditioning circuit includes a preamplifier circuit, a bandpass filter circuit, a post-amplifier circuit, an envelope extraction circuit, and a microcontroller main control circuit, wherein: The preamplifier circuit uses an instrumentation amplifier to form a differential amplifier circuit, which initially amplifies the microvolt-level ECG signal. The gain is set to about 100 times to improve the signal-to-noise ratio and suppress common-mode interference.

[0024] The bandpass filter circuit uses operational amplifiers to construct a passive filter, achieving high-pass filtering (cutoff frequency ~20Hz, filtering out low-frequency interference such as breathing sounds) and low-pass filtering (cutoff frequency ~4.8kHz, filtering out high-frequency electronic noise), which together form a bandpass filter that matches the effective frequency spectrum range of heart sounds.

[0025] The subsequent amplifier circuit is a non-inverting amplifier structure with a feedback resistor. R 1 =51 kΩ Grounding resistance R 2 =10 kΩ The gain is as follows:

[0026] Envelope extraction circuitry is used to assist in heart rhythm recognition by extracting contours representing the intensity of cardiac contraction and relaxation from the raw heart sound (or pulse wave) signal, thereby helping the algorithm to more easily identify the heart rhythm. The envelope extraction circuitry consists of precision rectifier diodes and an RC filter network, with the time constant designed using resistor R and capacitor C. The magnitude of the time constant in this invention. Designed as follows:

[0027] This time constant allows the output waveform to smoothly follow the peak energy changes of S1 and S2, while filtering out high-frequency details. The generated envelope signal is acquired by the ADC1 channel of the MCU chip and processed in conjunction with the original waveform acquired by the ADC0 channel to realize heart rate calculation and cycle segmentation.

[0028] The MCU chip of the microcontroller main control circuit is STM32F103C8T6. It acquires the original heart sound waveform after amplification through the ADC0 (PA0) port and the energy curve after envelope detection through the ADC1 (PA1) port. Subsequent processing through the software algorithm in step S2 can realize: heart rate cycle detection, S1 / S2 classification, heart rate calculation, murmur analysis, etc.

[0029] S2: Perform preprocessing and intelligent denoising on the digital signal, and extract bispectral features from the denoised clean digital signal. The process is as follows: Figure 4 As shown.

[0030] This step is achieved through a real-time heart sound signal intelligent denoising and feature extraction module. This module preprocesses and intelligently denoises the real-time digital signal and extracts bispectral features from the digital signal.

[0031] The digital signal preprocessing process is as follows:

[0032] Resampling: The input signal is resampled to 1000Hz to standardize the data.

[0033] Digital bandpass filtering: A finite-length impulse response (FIR) filter is used to perform secondary filtering in the 25~400Hz range to ensure signal purity.

[0034] Normalization: Normalize the signal amplitude to the range of [-1,1] to eliminate amplitude differences caused by different contact pressures during acquisition.

[0035] Heart sound segmentation: The energy envelope method is used to locate the cardiac cycle. The first heart sound (S1) and the second heart sound (S2) are located using a segmentation algorithm. Then, a signal segment containing at least two complete heartbeat cycles (approximately 2.5 seconds) is extracted. It is used for subsequent intelligent denoising and high-order spectral feature calculation and analysis.

[0036] The intelligent noise reduction process for real-time heart sound signals is as follows: In real-time heart sound denoising, the temporal segments of the continuously input heart sound stream are denoised. It is considered as a sequence arriving over time, and has a length of . L Step size is H The overlapping block segmentation process divides the continuous input data stream into overlapping blocks. k Each frame data block can be written as:

[0037] To adapt to the structural requirements of U-Net, the framed data blocks are first linearly resampled to the standard length of the intelligent denoising U-Net model within the local field of view, and then inversely resampled to restore the data after entering the network for processing, so that the intelligent denoising U-Net only needs to process inputs of fixed dimensions.

[0038] The tensors entering the network undergo transformations between the channel and time dimensions, and pass through multiple levels of convolution, downsampling, upsampling, and an attention module. The attention layer enhances the transient structure of heart sounds by implicitly modeling local correlations, while downsampling expands the effective receptive field, and upsampling is used to recover temporal details. The network outputs a denoised, clean heart sound signal stream, denoted as... After the corresponding inverse transformation and linear interpolation, the block is restored to its original length. Specifically: The forward inference of the intelligent denoising U-Net network model captures local context and outputs an estimate of background noise through multi-level convolution and attention mechanisms.

[0039] The intelligent denoising U-Net network model employs a one-dimensional U-Net architecture, achieving end-to-end denoising of heart sound signals through a symmetrical encoder-decoder structure. The encoder downsamples layer by layer to extract multi-scale features and introduces a self-attention mechanism after each resolution to model global dependencies. The decoder gradually recovers the signal length through upsampling, while simultaneously using skip connections to fuse the fine-grained details of the encoder with the semantic features of the decoder. This effectively suppresses noise while preserving the fine temporal structure of the signal, ultimately outputting a clean heart sound waveform. The structure of the intelligent denoising U-Net network model is attached. Figure 3 As shown, SA (SelfAttention) is the self-attention module. Down represents downsampling, Up represents upsampling, and skip (skip connections) refers to skip connections in a neural network. DoubleConv represents two consecutive convolutional layers (Conv1d or Conv2d), each followed by BatchNorm / GroupNorm and activation (ReLU / LeakyReLU), i.e., Conv→Norm→Act→Conv→Norm→Act. Such double convolution can stably train and increase non-linear expressive power.

[0040] In temporal modeling, it's used to capture long-range dependencies (non-local information). The input is (B, C, L), and Attention typically rearranges it to (B, L, C) to calculate the Query / Key / Value, or performs attention along the channel dimension. The core formula for Scaled dot-product attention is:

[0041] in They are obtained from the input through linear projection.

[0042] Attention allows the model to directly correlate locations that are far apart in time, which is useful for patterns in heart sounds that may span multiple time periods (e.g., periodic but deformed heart sounds). In the model of this invention, SA appears at multiple scales, injecting long-range dependency modeling capabilities at each key level of encoding and decoding: shallow layers can model detailed local interrelationships, while deep layers can model patterns across longer time scales.

[0043] The data processing procedure for the intelligent denoising U-Net network model is as follows: Input processing: The input tensor has the shape (batch, seq_len, in_c). First, the channel dimension is shifted to the second dimension using permute(0,2,1), becoming (batch, in_c, seq_len). Then, it is linearly interpolated to a fixed length of 2048 to adapt to the size changes inside the network.

[0044] The encoder (downsampling path) is as follows: Initial layer: DoubleConv(in_c,8) maps the number of input channels to 8 and extracts shallow features.

[0045] Then, the SA and Down modules are used alternately, for a total of four downsampling stages: SA(8)→Down(8,8), downsampling, the number of channels remains the same, but the length is halved; SA(8)→Down(8,16), the number of channels increases to 16; SA(16)→Down(16,32), the number of channels increases to 32; SA(32)→Down(32,64), the number of channels increases to 64.

[0046] Each Down module reduces the temporal resolution through convolution or pooling, while SA captures global dependencies.

[0047] Bottleneck layer: DoubleConv(64,64) further processes the deepest features.

[0048] The decoder (upsampling path) also has four upsampling stages. Each stage first upsamples through the Up module and fuses the skip connections, and then uses the SA module.

[0049] Finally, the channel is compressed to 1 using DoubleConv(8,1) to obtain the denoised signal.

[0050] Output processing: The result is linearly interpolated back to the original length of 2500, and then the shape (batch, seq_len, 1) is restored using permute(0,2,1).

[0051] The entire network achieves multi-scale feature extraction and reconstruction through a symmetrical encoder-decoder structure and skip connections, making it suitable for noise suppression of heart sound signals.

[0052] Adding an autoattention (SA) module after each downsampling and upsampling operation captures long-range dependencies: convolutional operations are limited by local receptive fields, while self-attention can model globally, enabling the network to focus on key components that are far apart in the heart sound signal (such as S1 and S2 heart sounds in the heart sound cycle), thus more accurately distinguishing signals from noise. Enhanced feature representation: Applying self-attention at different scales strengthens the contextual information of feature maps at each resolution, improving the model's robustness to noise morphology. For example, the feature map resolution decreases after downsampling, but self-attention preserves the global structure, avoiding the loss of details. Facilitated information flow: Adding self-attention after upsampling helps fuse details from skip connections with semantic information from the current layer, making the reconstructed signal smoother and more natural.

[0053] Then, preliminary signal purification is performed: the output of the intelligent denoising U-Net model after mapping back to the original length is used as the residual noise signal block. After removing noise, the time-domain heart sound signal is obtained after preliminary purification by the intelligent denoising U-Net. :

[0054] To further suppress background noise that fluctuates over time, online smoothing estimation and Wiener filtering are introduced. First, the signal is transformed to the frequency domain:

[0055] Online noise power spectrum estimation: Noise frequency domain representation estimated using intelligent denoising U-Net To smoothly update the background noise power spectrum:

[0056] in For the online noise power spectrum estimation after the k-th frame update, This is a smoothing coefficient, usually close to 1, such as 0.95.

[0057] Calculate the Wiener gain: Using the power spectrum of the preliminary purified signal and smoothed noise power spectrum Calculate each frequency point Gain coefficient on the input. Introducing hyperparameters. Controlling noise reduction intensity:

[0058] in : Approximate Wiener filter gain for the k-th frame.

[0059] Frequency domain amplitude correction: The gain is applied to the initially purified spectrum to complete the frequency domain filtering:

[0060] This represents the purified spectrum after frequency domain correction using Wiener filtering.

[0061] Inverse transform and overlap-add reconstruction: The frequency-corrected spectrum Returning to the time domain via Inverse Fourier Transform (IFFT):

[0062] This represents the cleaned signal block that has been converted back to the time domain after final correction.

[0063] Finally, the adjacent blocks are joined together using the Overlap-Add (OLA) method. Since the previous method used a step size of... H The sliding window requires each block to be placed back to its corresponding global timeline position and accumulated (usually in conjunction with a composite window function). To eliminate block boundary artifacts, a low-latency, continuous, and consistent global pure heart sound signal is obtained. :

[0064] Blocks are concatenated using an overlapping and additive approach to ensure a continuous and consistent overall output. The transient integrity of heart sound events is maintained through the model's local context structure and spectral domain gain control. The entire process involves block-by-block inference as the signal arrives, thus enabling stable real-time streaming heart sound denoising with extremely low latency.

[0065] The core feature extraction process based on bispectral density is as follows: Traditional heart sound analysis often relies on linear features such as time-domain envelope, power spectrum, or Mel spectrum. However, valve opening and closing, turbulent blood flow, and ventricular wall vibration all exhibit significant nonlinear interactions, containing rich quadratic phase coupling (QPC) information. Therefore, bispectrum is introduced as a core feature to capture the nonlinear structure revealed by third-order statistics.

[0066] The steps for bispectral estimation are as follows: Given a preprocessed heart sound sequence ,Will Frames, Frame Length ,overlapping .

[0067] Perform a Fast Fourier Transform (FFT) on each frame:

[0068] Calculate the frame-by-frame third-order spectrum:

[0069] Global bispectral density is obtained by averaging multiple frames:

[0070] The final result is a 256×256 two-dimensional bispectral matrix. This includes key nonlinear features of heart sound signals.

[0071] The amplitude values ​​of the bispectral matrix are linearly normalized to [0, 255], and then reshaped into a two-dimensional image tensor of size 1×256×256, which is used as input to a deep learning classification model. This process structurally enhances the one-dimensional heart sound sequence into a two-dimensional spectrogram with spatial patterns, which is beneficial for convolutional networks to extract higher-order features.

[0072] S3: Input the bispectral feature two-dimensional image tensor into the trained heart sound intelligent classification network model, and output the probability distribution of five states: normal heart sound, aortic stenosis, mitral regurgitation, mitral stenosis and mitral prolapse.

[0073] This step is achieved through the intelligent classification and diagnosis module.

[0074] The classification task is handled by a specially designed lightweight convolutional neural network. For example... Figure 5 As shown, the intelligent heart sound classification network model structure includes two cascaded convolutional blocks. The first convolutional block is configured with 32 3×3 convolutional kernels, and the second convolutional block is configured with 64 3×3 convolutional kernels. Each convolutional layer is followed by a ReLU activation function and a 2×2 max pooling layer to extract texture and shape features from the bispectral map. The extracted high-dimensional feature vectors are input to a fully connected layer containing 128 neurons, and finally output through a Softmax layer the probability distributions for five states: normal heart sound, aortic stenosis (AS), mitral regurgitation (MR), mitral stenosis (MS), and mitral valve prolapse (MVP).

[0075] The network structure is as follows: Input layer: Bispectral (1×256×256) Convolutional layer 1: 32 3×3 convolutional kernels, ReLU activation, 2×2 max pooling Convolutional layer 2: 64 3×3 convolutional kernels, ReLU activation, 2×2 max pooling Fully connected layer: 128-dimensional hidden units Output layer: Softmax, 5 classes in total.

[0076] The model was trained as follows: Framework: PyTorch.

[0077] Data source: PhysioNet CinC Challenge 2016 and samples from partner hospitals, totaling >3000 cases.

[0078] Optimizer: Adam, learning rate (1 × 10^-4).

[0079] Loss function: Cross-entropy loss.

[0080] Training epochs: 1500 epochs.

[0081] Figure 6 This demonstrates a comparison of the time-domain waveforms of the same heart sound signal before and after noise reduction.

[0082] The present invention has been described in detail above through embodiments, but the content described is only an exemplary embodiment of the present invention and should not be considered as limiting the scope of the present invention. The scope of protection of the present invention is defined by the claims. Any technical solutions designed by those skilled in the art using the technical solutions described in the present invention, or similar technical solutions designed by those skilled in the art under the inspiration of the technical solutions of the present invention, within the substance and scope of protection of the present invention, to achieve the above-mentioned technical effects, or equivalent changes and improvements made to the scope of the application, should still fall within the patent protection scope of the present invention. It should be noted that, for clarity, descriptions of some components and processes that are not directly and obviously related to the scope of protection of the present invention but are known to those skilled in the art have been omitted in the description of the present invention.

Claims

1. A smart heart sound classification method based on real-time streaming signal denoising and bispectral features, characterized in that, Includes the following steps: S1: Collect heart sound signals through a heart sound sensor, amplify and filter the initial signal, and complete the envelope extraction and digital conversion of the heart sound signal; S2: Perform preprocessing and intelligent denoising on the digital signal, and extract bispectral features from the denoised clean digital signal; S3: Input the bispectral feature two-dimensional image tensor into the trained heart sound intelligent classification network model, and output the probability distribution of five states: normal heart sound, aortic stenosis, mitral regurgitation, mitral stenosis and mitral prolapse.

2. The method according to claim 1, characterized in that, In step S1, the initial heart sound signal is first amplified and bandpass filtered, then amplified further and digitized by an MCU chip. An envelope extraction circuit is set between the primary and secondary amplification circuits to set the time constant. .

3. The method according to claim 2, characterized in that, Envelope extraction circuit setting time constant It takes 103ms.

4. The method according to claim 1, characterized in that, The preprocessing in step S2 includes the following: First, resampling is performed to uniformly resample the input signal to 1000Hz; Secondly, digital bandpass filtering is performed, and secondary filtering is carried out in the range of 25~400Hz. Then, normalization is performed to normalize the signal amplitude to the range of [-1, 1]. Finally, heart sound segmentation is performed, and the energy envelope method is used to locate the cardiac cycle. The first and second heart sounds are located by segmentation algorithm. Then, signal segments containing at least two complete heartbeat cycles are extracted for subsequent intelligent denoising and high-order spectral feature calculation and analysis.

5. The method according to claim 1, characterized in that, The intelligent denoising process is as follows: The streaming heart sound signal is processed in frames, and each frame is input into the intelligent denoising U-Net network. The network estimates the background noise component in the frame and subtracts it from the original signal to obtain the preliminary cleaned signal. Then, the preliminary cleaned signal and its corresponding noise estimate are transformed to the frequency domain, the noise power spectrum is updated by recursive smoothing, and the Wiener gain is calculated to perform amplitude correction on the spectrum. Finally, the continuous time domain signal is reconstructed by inverse Fourier transform and overlapping addition method.

6. The method according to claim 5, characterized in that, The intelligent denoising U-Net network model structure is as follows: This network adopts a one-dimensional U-Net architecture, which realizes end-to-end heart sound signal denoising through a symmetrical encoder-decoder structure. The encoder downsamples layer by layer to extract multi-scale features, and introduces a self-attention mechanism after each resolution to model global dependencies. The decoder gradually recovers the signal length through upsampling, and at the same time uses skip connections to fuse the fine-grained details of the encoder with the semantic features of the decoder, thereby effectively suppressing noise while preserving the fine time structure of the signal, and finally outputting a clean heart sound waveform.

7. The method according to claim 1, characterized in that, During bispectral feature extraction, the preprocessed continuous time-domain heart sound signal is divided into frames, with a frame length of [missing information]. ,overlapping This yields a 256×256 two-dimensional bispectral matrix. The amplitude values ​​of the bispectral matrix are linearly normalized to [0, 255], and then reshaped into a two-dimensional image tensor of size 1×256×256, which is used as the input of the deep learning classification model.

8. The method according to claim 7, characterized in that, The structure of the intelligent heart sound classification network model is as follows: It contains two cascaded convolutional blocks. The first convolutional block is configured with 32 3×3 convolutional kernels, and the second convolutional block is configured with 64 3×3 convolutional kernels. Each convolutional layer is followed by a ReLU activation function and a 2×2 max pooling layer to extract texture and shape features from the bispectral map. The extracted high-dimensional feature vector is input to a fully connected layer containing 128 neurons. Finally, the probability distribution of five states—normal heart sound, aortic stenosis, mitral regurgitation, mitral stenosis, and mitral valve prolapse—is output through a Softmax layer.

9. A smart heart sound classification device based on real-time streaming signal denoising and high-order spectral feature extraction, characterized in that, include: The heart sound sensing and acquisition module is used to acquire the initial heart sound signal and complete the digital conversion; The real-time heart sound signal intelligent denoising and feature extraction module uses a built-in algorithm to denoise the real-time heart sound digital signal and extract the bispectral features of the denoised heart sound digital signal; The intelligent classification and diagnosis module classifies the input bispectral features using a trained deep learning intelligent classification model and outputs the probability distribution of five states: normal heart sounds, aortic stenosis, mitral regurgitation, mitral stenosis, and mitral valve prolapse.

10. The apparatus according to claim 9, characterized in that, By connecting to an intelligent language model platform, it interacts with users and provides personalized suggestions based on their interactions.

Images