Blood pressure detection and classification system and method based on neural network, medium and terminal

Abstract

The invention provides a blood pressure detection and classification system and method based on a neural network, a medium and a terminal. Mel-frequency cepstral coefficient feature extraction, first heart sound time domain feature extraction and second heart sound time domain feature extraction are performed on acquired heart sound signal real-time data; the extracted Mel frequency cepstral coefficient feature data, the first heart sound time domain feature data and the second heart sound time domain feature data are used as model features to be input, time domain features and frequency domain features are included, and comprehensive coverage of time domain information and frequency domain information is achieved. The multi-input time domain convolutional neural network model is used for processing the three feature inputs and fusing the three feature inputs, the current blood pressure state of the user can be accurately detected, convenience and rapidness are achieved, and the result is accurate. The algorithm for extracting the frame where the heart sound peak point is located based on the MFCC0 is achieved, the influence of noise can be effectively avoided, the algorithm is simpler and more accurate, meanwhile, higher fault tolerance is achieved, and the first heart sound and the second heart sound can be accurately distinguished.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a blood pressure detection and classification system, method, medium and terminal based on neural networks. Background Technology

[0002] With the accelerating pace of modern life and the increasing aging population, cardiovascular disease has become a major public health issue worldwide. Blood pressure, as a crucial indicator for assessing cardiovascular health, is vital for continuous, non-invasive, and convenient monitoring, which is essential for disease prevention, diagnosis, and evaluation of treatment effectiveness. However, traditional blood pressure measurement methods, such as electronic blood pressure monitors based on inflatable cuffs, while meeting clinical needs to some extent, suffer from limited availability, inconvenient operation, and the inability to achieve continuous monitoring, thus failing to meet the demands of modern medical care and personal health management.

[0003] In recent years, with the rapid development of sensor technology, signal processing technology, and artificial intelligence algorithms, blood pressure measurement technology based on non-contact biosignal monitoring has gradually emerged. Among them, heart sound signals, as a type of biosignal containing rich cardiovascular physiological information, have become a research hotspot in the field of continuous blood pressure monitoring due to their non-invasiveness, ease of acquisition, and close relationship with blood pressure.

[0004] Currently, blood pressure measurement technology based on heart sound signals has made some progress, but many challenges and shortcomings still exist. On the one hand, existing technologies mostly rely on complex data acquisition systems, such as multimodal data acquisition schemes that combine smartphone microphones, flashlights, and cameras. Although this achieves indirect blood pressure monitoring to some extent, the acquisition process is complex, the acquisition is difficult, it requires a high level of expertise from both the equipment and the operator, it is inconvenient to operate, and the prediction results only show the trend of blood pressure changes, making it impossible to intuitively understand whether the current blood pressure is within the normal range. Practicality and user experience urgently need to be improved.

[0005] On the other hand, some studies have attempted to achieve continuous non-invasive blood pressure monitoring by constructing blood pressure regression models based on heart sound signals. However, these methods often rely on specific user datasets for model training, resulting in poor user adaptability and insufficient model generalization ability. Furthermore, because the features used are relatively limited (e.g., only the second heart sound apex is used as a feature), the rich information in the heart sound signals is not fully explored, making the measurement results susceptible to noise interference and requiring improvement in accuracy. Summary of the Invention

[0006] In view of the shortcomings of the prior art described above, the purpose of this application is to provide a blood pressure detection classification system, method, medium and terminal based on neural networks, which can solve the technical problems existing in the current blood pressure detection technology, such as complex data acquisition process, high acquisition difficulty, high requirements for equipment and collector professionalism, inconvenient operation, and prediction results that only show the trend of blood pressure change, as well as poor user adaptability and insufficient generalization ability of the model.

[0007] To achieve the above and other related objectives, a first aspect of this application provides a blood pressure detection and classification system based on a neural network, comprising: a preprocessing module for preprocessing acquired real-time heart sound signal data; a first feature extraction module for extracting features from the preprocessed real-time heart sound signal data to generate a two-dimensional feature array of Mel-frequency cepstral coefficients; a second feature extraction module for generating a first two-dimensional time-domain feature array and a second two-dimensional time-domain feature array of heart sounds based on the Mel-frequency cepstral coefficients; and a blood pressure detection and classification module for generating a blood pressure detection and classification result based on the Mel-frequency cepstral coefficients, the first two-dimensional time-domain feature array of heart sounds, and the second two-dimensional time-domain feature array of heart sounds, and on a pre-constructed blood pressure detection and classification model, and outputting the classification result.

[0008] In some embodiments of the first aspect of this application, the second feature extraction module generates a first two-dimensional feature array and a second two-dimensional feature array of heart sounds in the time domain based on the two-dimensional feature array of Mel frequency cepstral coefficients, including: sequentially traversing the first row of feature vectors in the two-dimensional feature array of Mel frequency cepstral coefficients according to a preset window length to filter out all maximum points in the first row of feature vectors, and recording the signal frame where each maximum point is located and the corresponding energy value; filtering out all peak points and all low peak points in the first row of feature vectors according to a preset high energy threshold, a preset low energy threshold and the energy value corresponding to each maximum point, and recording the signal frame where each peak point is located and the corresponding energy value, and the signal frame where each low peak point is located and the corresponding energy value; calculating the number of first frames in which each peak point satisfies the preset high energy threshold within a preset value range, and calculating the number of second frames in which each low peak point satisfies the preset low energy threshold within a preset value range; determining the sum of the first frame count and the sum of the second frame count, and generating the first two-dimensional feature array of heart sounds in the time domain and the second two-dimensional feature array of heart sounds in the time domain based on the determination result.

[0009] In some embodiments of the first aspect of this application, the blood pressure detection classification module generates a blood pressure detection classification result and outputs the classification result based on the Mel frequency cepstral coefficient two-dimensional feature array, the first heart sound time-domain two-dimensional feature array, and the second heart sound time-domain two-dimensional feature array, and a pre-built blood pressure detection classification model. This includes: performing dimensionality transformation on the Mel frequency cepstral coefficient two-dimensional feature array, the first heart sound time-domain two-dimensional feature array, and the second heart sound time-domain two-dimensional feature array to generate a multi-channel Mel frequency cepstral coefficient one-dimensional time-domain feature array, a multi-channel first heart sound time-domain one-dimensional feature array, and a multi-channel second heart sound time-domain one-dimensional feature array; inputting the multi-channel Mel frequency cepstral coefficient one-dimensional time-domain feature array, the multi-channel first heart sound time-domain one-dimensional feature array, and the multi-channel second heart sound time-domain one-dimensional feature array into the pre-built blood pressure detection classification model to generate a blood pressure detection classification result and output the classification result.

[0010] In some embodiments of the first aspect of this application, a model building module is further included. This model building module is used to build a blood pressure detection classification model, comprising: preprocessing the acquired historical heart sound signal dataset; extracting features from the preprocessed historical heart sound signal dataset to generate a two-dimensional feature history array set of Mel-frequency cepstral coefficients; generating a first two-dimensional temporal feature history array set of heart sounds and a second two-dimensional temporal feature history array set of heart sounds based on the Mel-frequency cepstral coefficients two-dimensional feature history array set of heart sounds; the Mel-frequency cepstral coefficients two-dimensional feature history array set of heart sounds, the first two-dimensional temporal feature history array set of heart sounds, and the second two-dimensional temporal feature history array set of heart sounds constituting a heart sound signal training dataset; and training a multi-input temporal convolutional neural network model using the heart sound signal training dataset to build the blood pressure detection classification model.

[0011] In some embodiments of the first aspect of this application, the multi-input temporal convolutional neural network model includes three temporal convolutional neural networks, a vector concatenation layer, and a fully connected layer; each temporal convolutional neural network is merged through the vector concatenation layer, and the one-dimensional feature arrays output by each temporal convolutional neural network are concatenated through the vector concatenation layer and then output to the fully connected layer for calculation.

[0012] In some embodiments of the first aspect of this application, the temporal convolutional neural network includes: a first convolutional layer, a first normalized activation layer, a temporal convolutional layer, and a global average pooling layer; the temporal convolutional layer includes a plurality of temporal convolutional blocks; the temporal convolutional blocks are composed of a combination of a second convolutional layer and a second normalized activation layer; the first convolutional layer and the second convolutional layer are both one-dimensional convolutional layers.

[0013] In some embodiments of the first aspect of this application, the method by which the first feature extraction module extracts features from the preprocessed real-time heart sound signal data to generate a two-dimensional feature array of Mel-frequency cepstral coefficients includes: dividing the preprocessed real-time heart sound signal data into several signal frames according to a preset frame length and a preset frame shift; performing a weighting operation on each segmented signal frame according to a preset window function; performing a fast Fourier transform on each signal frame after the weighting operation to obtain the spectrum data corresponding to each signal frame; performing filtering processing on the spectrum data corresponding to each signal frame to obtain the filtering coefficients corresponding to each signal frame; performing logarithmic processing on the filtering coefficients corresponding to each signal frame, and performing discrete cosine transform processing on the filtering coefficients corresponding to each signal frame after the logarithmic processing to generate a two-dimensional feature array of Mel-frequency cepstral coefficients.

[0014] To achieve the above and other related objectives, a second aspect of this application provides a blood pressure detection classification method based on a neural network, comprising: preprocessing acquired real-time heart sound signal data; extracting features from the preprocessed real-time heart sound signal data to generate a two-dimensional feature array of Mel frequency cepstral coefficients; generating a first two-dimensional feature array of heart sound time domain and a second two-dimensional feature array of heart sound time domain based on the Mel frequency cepstral coefficients two-dimensional feature array; generating a blood pressure detection classification result based on the Mel frequency cepstral coefficients two-dimensional feature array, the first two-dimensional feature array of heart sound time domain and the second two-dimensional feature array of heart sound time domain, and a pre-constructed blood pressure detection classification model, and outputting the classification result.

[0015] To achieve the above and other related objectives, a third aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the neural network-based blood pressure detection and classification method as described above.

[0016] To achieve the above and other related objectives, a fourth aspect of this application provides an electronic terminal, including a memory, a processor, and a computer program stored in the memory; the processor executes the computer program to implement the neural network-based blood pressure detection and classification method as described above.

[0017] As described above, the neural network-based blood pressure detection and classification system, method, medium, and terminal of this application have the following advantages:

[0018] Beneficial effects:

[0019] (1) The blood pressure detection classification system based on neural network in this application is a portable blood pressure detection system based on a multi-input temporal convolutional neural network model. Compared with traditional blood pressure monitoring methods, it does not require going to a professional measurement institution or purchasing blood pressure measurement equipment. It has low requirements for the professionalism of the equipment and the collector. It only requires a mobile phone microphone to collect real-time heart sound signal data and perform blood pressure detection. The collection process is simple, the collection difficulty is low, and it is more convenient and faster.

[0020] (2) This application performs Mel-frequency cepstral coefficient (MFCC) feature extraction, first heart sound time-domain feature extraction, and second heart sound time-domain feature extraction on the collected real-time heart sound signal data. The extracted Mel-frequency cepstral coefficient (MFCC) feature data, first heart sound (S1) time-domain feature data, and second heart sound (S2) time-domain feature data are used as model feature inputs, including time-domain and frequency-domain features. Furthermore, the input features have a certain mapping relationship with systolic blood pressure, diastolic blood pressure, and mean blood pressure. Compared to traditional methods that often use only one time-domain or frequency-domain feature, this system uses more comprehensive feature inputs, simultaneously capturing the frequency components of the heart sound signal and the specific time points of cardiac activity, achieving comprehensive coverage of time-frequency domain information. By using a multi-input time-domain convolutional neural network model to process and fuse these three feature inputs, the system can accurately detect the user's current blood pressure status, which is convenient, fast, and accurate.

[0021] (3) This application implements a frame extraction algorithm for heart sound peak points based on MFCC_0. The heart sound peak point extraction algorithm is based on the first row feature vector (MFCC_0) in the two-dimensional feature array of Mel frequency cepstral coefficients (MFCC). MFCC_0 represents the total energy of the audio frame. According to the preset high energy threshold and the preset low energy threshold, the number of first frames in which each peak point meets the preset high energy threshold within the preset value range is calculated, and the number of second frames in which each low peak point meets the preset low energy threshold within the preset value range is calculated. The sum of the first frame number and the sum of the second frame number are judged. Based on the judgment result, the signal frame in which the first heart sound (S1) peak point and the signal frame in which the second heart sound (S2) peak point are located in the heart sound signal can be accurately identified, so as to generate the first heart sound time domain two-dimensional feature array and the second heart sound time domain two-dimensional feature array and use them as the feature input of the model. Compared with traditional heart sound peak extraction algorithms, the heart sound peak point extraction algorithm of this application does not need to find a specific point, but only needs to find the signal frame where the peak point is located. It can effectively avoid the influence of noise, is simpler and more accurate, has stronger fault tolerance, and can accurately distinguish between the first heart sound (S1) and the second heart sound (S2).

[0022] (4) This application constructs a blood pressure detection classification model by training a multi-input temporal convolutional neural network model, which has stronger learning and generalization capabilities, thus obtaining more accurate blood pressure classification results. It is applicable to any population and does not require pre-training based on user data for accurate and rapid blood pressure detection. Furthermore, compared to traditional neural network models, the multi-input temporal convolutional neural network model (MTCNN) proposed in this application simultaneously uses Mel-frequency cepstral coefficients (MFCC) feature data, first heart sound (S1) temporal feature data, and second heart sound (S2) temporal feature data as model inputs. This allows for a more comprehensive learning of the blood pressure-related features in the heart sound signal, resulting in stronger robustness and generalization capabilities. In addition, the multi-input temporal convolutional neural network model (MTCNN) uses a large number of one-dimensional temporal convolutions instead of traditional two-dimensional convolutions, paying more attention to temporal information and thus better learning the temporal features of the heart sound signal. It also compresses frequency domain information, reducing computational load. Ultimately, the multi-input temporal convolutional neural network model (MTCNN) proposed in this application can efficiently and accurately complete the blood pressure classification task with extremely low parameter and computational requirements. Attached Figure Description

[0023] Figure 1 The diagram shown is a schematic representation of a neural network-based blood pressure detection and classification system according to an embodiment of this application.

[0024] Figure 2 The diagram shown is a flowchart illustrating the extraction of Mel frequency cepstral coefficient features in one embodiment of this application.

[0025] Figure 3 The diagram shown is a flowchart illustrating the extraction of temporal features of heart sounds in one embodiment of this application.

[0026] Figure 4 The diagram shown is a schematic representation of the structure of a multi-input temporal convolutional neural network model in one embodiment of this application.

[0027] Figure 5 The diagram shown is a schematic representation of the structure of a temporal convolutional neural network in one embodiment of this application.

[0028] Figure 6 The diagram shown is a flowchart of a neural network-based blood pressure detection and classification method in one embodiment of this application.

[0029] Figure 7 The diagram shown is a structural schematic of an electronic terminal according to an embodiment of this application. Detailed Implementation

[0030] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, unless otherwise specified, the following embodiments and features in the embodiments can be combined with each other.

[0031] Before providing a further detailed description of this application, the nouns and terms used in the embodiments of this application are explained, and the nouns and terms used in the embodiments of this application shall be interpreted as follows:

[0032] <1> MFCC (Mel Frequency Cepstral Coefficients) are a type of feature widely used in speech and audio processing. They are based on the non-linear frequency perception of the human auditory system. MFCC features are particularly suitable for tasks such as speech recognition and speaker recognition because they can capture features in audio signals that are closely related to human auditory perception.

[0033] <2> Frame length: The number of sample points contained in each frame during the framing process.

[0034] <3> Frame shift: In signal processing, especially in the framing process of audio or speech signals, it refers to the number of sample points between the start points of two adjacent frames.

[0035] <4> The Gibbs effect is an important phenomenon in signal processing, especially noticeable when performing Fourier series expansions or Fast Fourier Transform (FFT). It refers to the oscillation phenomenon near discontinuities in a signal (such as jump points or step points), which cannot be completely eliminated even by using an infinite number of terms in a Fourier series expansion.

[0036] <5> Mel filter: A Mel filter is a type of filter based on the characteristics of sound perceived by the human ear. It is commonly used in tasks such as speech recognition, speech synthesis, and audio processing. Its design is based on the frequency response characteristics of the human auditory system to sound, and it can mimic the way the human ear perceives sound.

[0037] <6> Adam optimizer (Adaptive Moment Estimation) is a widely used deep learning optimization algorithm that aims to adjust the learning rate of each parameter by calculating the first and second moment estimates of the gradient, thereby achieving more efficient network training.

[0038] The technical solutions of this application can be applied to the field of artificial intelligence technology, particularly relating to a blood pressure detection and classification system, method, medium, and terminal based on a neural network. Regular blood pressure monitoring helps prevent cardiovascular disease; however, currently used blood pressure measurement methods often rely on inflatable cuffs, which are not widely available and lack portability, making it impossible to conveniently and quickly monitor blood pressure at any time. Heart sounds are audible sounds produced by the contraction and relaxation of the atria and ventricles, valvular movement, and blood flow-related dynamic events. Heart sounds mainly include the first heart sound (S1) and the second heart sound (S2). The first heart sound S1 marks the beginning of cardiac contraction, while the second heart sound S2 marks the end of ventricular contraction and the beginning of ventricular relaxation. Studies have shown that the time-frequency domain variations of heart sound signals, the duration and amplitude of the first heart sound S1 and the second heart sound S2, all have a certain mapping relationship with systolic blood pressure, diastolic blood pressure, and mean blood pressure. Therefore, reasonable analysis and processing of heart sound signals can effectively detect blood pressure. With its powerful performance, neural networks, after reasonable structural improvements and optimizations, can simultaneously process multiple features of heart sound signals and output calculation results to determine whether blood pressure is normal.

[0039] To address the problems mentioned above, this invention provides a neural network-based blood pressure detection and classification system, method, medium, and terminal. This addresses the technical issues in existing blood pressure detection technologies, such as complex data acquisition processes, high acquisition difficulty, high requirements for equipment and operator expertise, inconvenient operation, prediction results that only show blood pressure change trends, and poor user adaptability and insufficient generalization ability of the models.

[0040] To facilitate understanding of the embodiments of this application, firstly, in conjunction with Figure 1 Detailed explanation. Figure 1 A schematic diagram of the structure of a neural network-based blood pressure detection and classification system in this embodiment is shown. The neural network-based blood pressure detection and classification system 100 in this embodiment mainly includes: a preprocessing module 101, a first feature extraction module 102, a second feature extraction module 103, and a blood pressure detection and classification module 104. Wherein:

[0041] The preprocessing module 101 is used to preprocess the acquired real-time heart sound signal data.

[0042] The first feature extraction module 102 is used to extract features from the preprocessed real-time heart sound signal data and generate a two-dimensional feature array of Mel frequency cepstral coefficients.

[0043] The second feature extraction module 103 is used to generate a first two-dimensional feature array of heart sounds in the time domain and a second two-dimensional feature array of heart sounds in the time domain based on the two-dimensional feature array of Mel frequency cepstral coefficients.

[0044] The blood pressure detection classification module 104 is used to generate a blood pressure detection classification result and output the classification result based on the two-dimensional feature array of Mel frequency cepstral coefficients, the two-dimensional feature array of the first heart sound in the time domain and the two-dimensional feature array of the second heart sound in the time domain, and based on a pre-built blood pressure detection classification model.

[0045] In this embodiment, the method for acquiring the real-time heart sound signal data includes: using a mobile phone microphone to collect the real-time heart sound signal data. Specifically, in a quiet environment, the mobile phone microphone is used to record a period of time close to the heart, the recording time including but not limited to: 10s, 15s, 20s, 25s, etc., and the real-time heart sound signal data collected by the mobile phone microphone is transmitted to the preprocessing module 101.

[0046] In this embodiment, the preprocessing module 101 is used to preprocess the acquired real-time heart sound signal data. The preprocessing methods include downsampling and filtering.

[0047] In this embodiment, the downsampling process is the process of resampling the original signal at a lower frequency in time. Specifically, the original real-time heart sound signal data is downsampled to a preset frequency, which includes, but is not limited to, 2000Hz, 4000Hz, and 8000Hz. The purpose of downsampling is to reduce the amount of data and processing complexity, while retaining information in the real-time heart sound signal data that is useful for subsequent analysis. For real-time heart sound signal data, as a preferred embodiment, a sampling frequency of 4000Hz is usually sufficient to capture important cardiac activity information, while reducing the burden of storage and processing.

[0048] In this embodiment, the filtering process is a signal processing technique used to remove unwanted frequency components from a signal. Specifically, a Butterworth bandpass filter of 20Hz-1000Hz is used. A Butterworth filter is a filter with a flat frequency response; it has maximum flatness in the passband and gradually attenuates in the stopband. The Butterworth bandpass filter allows signals within a specific frequency range to pass through while attenuating signals at other frequencies. For real-time heart sound signal data, low-frequency noise below 20Hz may include breath sounds, ambient noise, etc., while high-frequency noise above 1000Hz may include electronic noise or unwanted signal interference. Therefore, using a Butterworth bandpass filter of 20Hz-1000Hz can effectively eliminate these noises and improve the signal-to-noise ratio.

[0049] In this embodiment, as Figure 2The diagram illustrates the flowchart for extracting Mel-frequency cepstral coefficient features in an embodiment of this application. The first feature extraction module 102 extracts features from the preprocessed real-time heart sound signal data, generating a two-dimensional feature array of Mel-frequency cepstral coefficients, including the following methods:

[0050] S201: Based on the preset frame length and preset frame shift, the preprocessed real-time heart sound signal data is divided into several signal frames.

[0051] In this embodiment, since the audio signal changes over time, it needs to be divided into a series of short-duration signal frames for analysis. Based on a preset frame length and a preset frame shift, the preprocessed real-time heart sound signal data is segmented into several signal frames. For example, 15 seconds of real-time heart sound signal data (60,000 points in total) with a sampling rate of 4000Hz is segmented into frames. The preset frame length is 256 points (corresponding to a time length of 256 / 4000Hz = 64ms), and the preset frame shift is 128 points (corresponding to a time length of 32ms), where 256 and 128 represent the number of samples. Frame segmentation divides the long-term varying audio signal into multiple short-duration stable signal segments (frames), making the signal characteristics within each frame relatively stable, facilitating subsequent processing. The preset frame shift ensures overlap between adjacent frames, thereby maintaining signal continuity and reducing the impact of discontinuities at frame boundaries on feature extraction.

[0052] S202: Perform weighted operations on each segmented signal frame according to the preset window function.

[0053] In this embodiment, the preset window function is a Hamming window. A Hamming window is applied to each segmented signal frame for weighting. The sample points of each signal frame are multiplied by the coefficients of the preset window function, thereby smoothing the edges of the signal frames, reducing spectral leakage, increasing the continuity of the signal at both ends of the frame, and reducing the Gibbs effect. The weighting operation causes the signal to gradually transition to zero at the boundaries, thus reducing spectral leakage and improving the accuracy of spectral analysis. The Hamming window is a commonly used window function that reduces the amplitude of the side lobes while maintaining the main lobe width, resulting in a smoother spectrum.

[0054] S203: Perform Fast Fourier Transform on each signal frame after weighting to obtain the spectral data corresponding to each signal frame.

[0055] In this embodiment, each signal frame after weighting is processed by a Fast Fourier Transform (FFT) to transform it from the time domain to the frequency domain, resulting in a frequency domain representation of each signal frame, i.e., spectral data. This spectral data describes the amplitude and phase information of different frequency components in the signal frame. Specifically, the Fast Fourier Transform (FFT) converts the signal frame in the time domain (the weighted signal frame) into a complex sequence in the frequency domain, where each complex element corresponds to a frequency component. The modulus (or amplitude) of this complex sequence represents the intensity or energy of the frequency component, while the phase represents the phase difference of the frequency component relative to a reference point. The result of the Fast Fourier Transform (FFT) is typically represented as a series of amplitude values ​​(and sometimes phase values) at frequency points uniformly distributed across the frequency range covered by the signal frame. These amplitude values ​​can be used for further analysis, such as calculating power spectral density and performing spectral analysis. In speech signal processing, the Fast Fourier Transform (FFT) is one of the key steps in extracting spectral features. By performing Fast Fourier Transform (FFT) processing on each signal frame, the spectral representation of that frame can be obtained, and features such as MFCC (Mel frequency cepstral coefficients) can be extracted for subsequent tasks such as speech recognition and speech synthesis.

[0056] S204: Filter the spectral data corresponding to each signal frame to obtain the filter coefficients corresponding to each signal frame.

[0057] In this embodiment, 20 Mel filters are used to filter the spectral data after Fast Fourier Transform (FFT) processing, resulting in a set of Mel filter coefficients that reflect the energy distribution of the signal frame across different frequency bands on the Mel frequency scale. The Mel filter bank is designed based on the auditory characteristics of the human ear, enabling a better simulation of human sound perception. Mel filtering reduces the dimensionality of the spectrum while preserving important information in the speech signal.

[0058] In this embodiment, the Mel filter bank is implemented by dividing the linear frequency into several triangular filters, each corresponding to a Mel frequency band. The Mel frequency scale is based on human auditory perception and its relationship with the linear frequency is non-linear: in the low-frequency region, the Mel frequency and the linear frequency are approximately linearly related; while in the high-frequency region, the Mel frequency and the linear frequency have a logarithmic relationship. This design allows the Mel filter bank to more accurately reflect the human ear's perception of sound. The Mel filter bank typically includes a series of triangular filters, usually between 20 and 40. The center frequency and bandwidth of each filter are determined according to the Mel scale. The frequency response function of these filters is a triangle, reaching its maximum value at the center frequency, and then gradually decreasing to the left and right until the frequency reaches 0.

[0059] S205: Take the logarithm of the filter coefficients corresponding to each signal frame, and perform discrete cosine transform on the filter coefficients corresponding to each signal frame after logarithmic processing to generate a two-dimensional feature array of Mel frequency cepstral coefficients.

[0060] In this embodiment, the filter coefficients corresponding to each signal frame are logarithmically calculated to simulate the logarithmic perception characteristic of sound by the human ear. For example, the base-2 logarithm (log2) is taken for each filter coefficient corresponding to each signal frame. This logarithmic operation compresses the dynamic range of the data, resulting in a more uniform data distribution, which is beneficial for subsequent processing.

[0061] In this embodiment, the filter coefficients corresponding to each signal frame after logarithmic processing are subjected to Discrete Cosine Transform (DCT) to obtain a two-dimensional feature array of Mel-frequency cepstral coefficients (MFCC features). These features are commonly used as acoustic features for classification and modeling in tasks such as speech recognition. The Discrete Cosine Transform (DCT) process has the characteristic of energy concentration, which can concentrate the main energy of the signal in the low-frequency part, which is beneficial to subsequent feature extraction and recognition. Through Discrete Cosine Transform (DCT) processing, the dimensionality of the features can be further reduced, thereby improving recognition efficiency.

[0062] In this embodiment, by sequentially performing frame segmentation, weighting, fast Fourier transform, filtering, logarithmic transformation, and discrete cosine transform on the preprocessed real-time heart sound signal data, the 15-second real-time heart sound signal data (60,000 points in total) with a sampling rate of 4000Hz can be transformed into two-dimensional 20×468 MFCC features, where 20 represents the number of Mel filters and 468 represents the number of frames. This completes the Mel frequency cepstral coefficient (MFCC) feature extraction.

[0063] In this embodiment, as Figure 3 The diagram illustrates the process of extracting time-domain features of heart sounds in an embodiment of this application. The second feature extraction module 103 generates the first and second two-dimensional time-domain feature arrays of heart sounds based on the Mel frequency cepstral coefficients two-dimensional feature array using a frame extraction algorithm for the heart sound peak points, including the following steps S301-S304:

[0064] S301: According to the preset window length, traverse the first row of feature vectors in the two-dimensional feature array of Mel frequency cepstral coefficients in sequence to filter out all the maximum points in the first row of feature vectors, and record the signal frame where each maximum point is located and the corresponding energy value.

[0065] In this embodiment, the first row of the feature vector (MFCC_0) in the Mel Frequency Cepstral Coefficients (MFCC) two-dimensional feature array contains short-time energy information of the real-time heart sound signal data; that is, the value of MFCC_0 represents the total energy or power of the audio frame. Therefore, MFCC_0 can be used to detect the signal frame where the heart sound peak point is located in the real-time heart sound signal data.

[0066] In this embodiment, each point in the first row of the feature vector (MFCC_0) represents the total energy of the signal frame. For example, if the Mel-frequency cepstral coefficients (MFCC) two-dimensional feature array is 20×468 feature data, then the first row of the feature vector (MFCC_0) is 1×468 one-dimensional data. A preset high-energy threshold and a preset low-energy threshold are used to detect the signal frames containing the peak points of the first and second heart sound signals, respectively.

[0067] In this embodiment, the preset window length is the length of 10 signal points (frames). Using the length of 10 signal points (frames) as a smoothing window, all maxima in the first row of feature vectors (MFCC_0) are calculated, and the signal frames containing each maxima and their corresponding energy values ​​are recorded. In the first row of feature vectors (MFCC_0), a maxima is a point that is locally the largest, meaning its value is greater than the values ​​of its neighboring points.

[0068] S302: Based on the preset high energy threshold, the preset low energy threshold, and the energy values ​​corresponding to each of the maximum points, filter out all high peak points and all low peak points in the first row of feature vectors, and record the signal frame and corresponding energy value of each of the high peak points, and the signal frame and corresponding energy value of each of the low peak points.

[0069] In this embodiment, the preset high energy threshold is the average energy of all signal points (frames) in the first row of the feature vector (MFCC_0), which is AVG; the preset low energy threshold is AVG / 2. In the first row of the feature vector (MFCC_0), the maximum value points that are greater than the preset high energy threshold (AVG) are peak points, and the maximum value points that are greater than the preset low energy threshold (AVG / 2) and less than the preset high energy threshold (AVG) are low peak points. The signal frames and corresponding energy values ​​of each peak point, and the signal frames and corresponding energy values ​​of each low peak point are recorded.

[0070] S303: Calculate the number of first frames in which each of the peak points satisfies the preset high energy threshold within a preset value range, and calculate the number of second frames in which each of the low peak points satisfies the preset low energy threshold within a preset value range.

[0071] In this embodiment, the preset value range is calculated by taking 15 points to the left and 15 points to the right of each of the peak points or each of the low peak points, respectively, and calculating the number of first frames where each of the peak points is greater than a preset high energy threshold within the preset value range, and calculating the number of second frames where each of the low peak points is greater than a preset low energy threshold and less than a preset high energy threshold within the preset value range.

[0072] S304: Determine the sum of the first frame count and the sum of the second frame count, and generate a first two-dimensional feature array of the heart sound in the time domain and a second two-dimensional feature array of the heart sound in the time domain based on the determination result.

[0073] In this embodiment, since the duration of the first heart sound (S1) is relatively long, approximately twice that of the second heart sound (S2), the first heart sound (S1) and the second heart sound (S2) can be distinguished based on the sum of the first frame count and the sum of the second frame count.

[0074] In this embodiment, if the sum of the first number of frames is greater than the sum of the second number of frames, then all signal frames containing the peak points selected in step S302 are signal frames containing the first heart sound peak points, and all signal frames containing the low peak points are signal frames containing the second heart sound peak points. A preset number of signal frames containing the peak points are selected from the first peak point to generate a first heart sound time-domain two-dimensional feature array, and a preset number of signal frames containing the low peak points are selected from the first low peak point to generate a second heart sound time-domain two-dimensional feature array.

[0075] In this embodiment, if the sum of the first number of frames is less than the sum of the second number of frames, then all signal frames containing low peak points selected in step S302 are signal frames containing the first heart sound peak point, and all signal frames containing high peak points are signal frames containing the second heart sound peak point. A preset number of signal frames containing low peak points are selected from the first low peak point to generate a first heart sound time-domain two-dimensional feature array, and a preset number of signal frames containing high peak points are selected from the first high peak point to generate a second heart sound time-domain two-dimensional feature array.

[0076] In this embodiment, for example, in the first row of feature vector (MFCC_0), if there are 50 maxima, of which 30 are peak points and 20 are trough points, 15 points are taken to the left and 15 points to the right of each peak point, and the number of first frames around each peak point that are greater than a preset high-energy threshold is calculated. All the calculated first frame numbers are then summed. Similarly, 15 points are taken to the left and 15 points to the right of each trough point, and the number of second frames around each trough point that are greater than a preset low-energy threshold and less than a preset high-energy threshold is calculated. All the calculated second frame numbers are then summed. For example, if the sum of the number of first frames surrounding 30 peak points that are greater than a preset high energy threshold is 200, and the sum of the number of second frames surrounding 20 low peak points that are greater than a preset low energy threshold and less than a preset high energy threshold is 100, then the signal frames containing these 30 peak points are all signal frames containing the first heart sound peak point, and the signal frames containing these 20 low peak points are all signal frames containing the second heart sound peak point. Then, among these 30 peak points, the signal frames containing the first 10 peak points are selected to generate a two-dimensional time-domain feature array for the first heart sound, and among these 20 low peak points, the signal frames containing the first 10 low peak points are selected to generate a two-dimensional time-domain feature array for the second heart sound. For example, if the Mel-frequency cepstral coefficients (MFCC) two-dimensional feature array is 20×468 two-dimensional frequency domain data, then the two-dimensional time-domain feature arrays for the first heart sound (S1) and the second heart sound (S1) are both 10×256 two-dimensional time-domain data.

[0077] In this embodiment, for example, in the first row of feature vector (MFCC_0), if there are 50 maxima, of which 30 are peak points and 20 are trough points, 15 points are taken to the left and 15 points to the right of each peak point, and the number of first frames around each peak point that are greater than a preset high-energy threshold is calculated. All the calculated first frame numbers are then summed. Similarly, 15 points are taken to the left and 15 points to the right of each trough point, and the number of second frames around each trough point that are greater than a preset low-energy threshold and less than a preset high-energy threshold is calculated. All the calculated second frame numbers are then summed. For example, if the sum of the number of first frames around the 30 peak points that are greater than the preset high energy threshold is 100, and the sum of the number of second frames around the 20 low peak points that are greater than the preset low energy threshold and less than the preset high energy threshold is 200, then the signal frames where these 20 low peak points are located are all signal frames where the first heart sound peak point is located, and the signal frames where these 30 peak points are located are all signal frames where the second heart sound peak point is located. Then, among these 20 low peak points, the signal frames where the first 10 low peak points are located are selected to generate the first heart sound time-domain two-dimensional feature array, and among these 30 high peak points, the signal frames where the first 10 high peak points are located are selected to generate the second heart sound time-domain two-dimensional feature array.

[0078] In this embodiment, the blood pressure detection classification module 104 generates a blood pressure detection classification result and outputs the classification result based on the two-dimensional feature array of Mel frequency cepstral coefficients, the two-dimensional feature array of the first heart sound time domain, and the two-dimensional feature array of the second heart sound time domain, and based on a pre-built blood pressure detection classification model.

[0079] (1) Perform dimension-changing processing on the two-dimensional feature array of Mel frequency cepstral coefficients, the two-dimensional feature array of the first heart sound in the time domain, and the two-dimensional feature array of the second heart sound in the time domain, respectively, to generate a one-dimensional feature array of multi-channel Mel frequency cepstral coefficients in the time domain, a one-dimensional feature array of multi-channel first heart sound in the time domain, and a one-dimensional feature array of multi-channel second heart sound in the time domain.

[0080] In this embodiment, before inputting the Mel frequency cepstral coefficient (MFCC) two-dimensional feature array, the first heart sound time-domain two-dimensional feature array, and the second heart sound time-domain two-dimensional feature array into the pre-constructed blood pressure detection classification model, a dimension-shifting process is performed to generate a multi-channel Mel frequency cepstral coefficient (MFCC) one-dimensional time-domain feature array, a multi-channel first heart sound time-domain one-dimensional feature array, and a multi-channel second heart sound time-domain one-dimensional feature array. For example, the Mel frequency cepstral coefficient (MFCC) two-dimensional feature array is 20×468 two-dimensional frequency domain data, and the first heart sound (S1) (S1) and second heart sound (S1) (S1) time-domain two-dimensional feature arrays are both 10×256 two-dimensional time-domain data. After the dimension-shifting process, the Mel frequency cepstral coefficient (MFCC) two-dimensional feature array becomes a 20×1×468 one-dimensional time-domain feature array, that is, the original single-channel two-dimensional frequency domain feature input is transformed into a multi-channel one-dimensional time-domain feature input. Similarly, the two-dimensional time-domain feature arrays of the first heart sound (S1) and the second heart sound (S1) are transformed into a one-dimensional time-domain feature array of 10×1×256. In this way, the one-dimensional convolutional kernels of the temporal convolutional neural network (TCNN) in the blood pressure detection classification model can be used to give more attention to the time-domain information while compressing the frequency-domain information, thereby reducing the size of the feature map and lowering the computational load.

[0081] (2) Input the one-dimensional time-domain feature array of the multi-channel Mel frequency cepstral coefficients, the one-dimensional time-domain feature array of the multi-channel first heart sound, and the one-dimensional time-domain feature array of the multi-channel second heart sound into the pre-constructed blood pressure detection classification model, generate blood pressure detection classification results, and output the classification results.

[0082] In this embodiment, the blood pressure detection classification result is a one-dimensional vector, specifically a 1×3 vector, representing three probability values, corresponding to the labels of normal blood pressure, high blood pressure, and low blood pressure, respectively. The label corresponding to the maximum value is taken as the final blood pressure detection classification result and displayed. For example, if the maximum value is the first value of the 1×3 vector, then the output is "normal blood pressure"; if the second value is the largest, then the output is "high blood pressure"; and if the third value is the largest, then the output is "low blood pressure".

[0083] In this embodiment, the neural network-based blood pressure detection classification system 100 further includes a model building module 105, which is used to build a blood pressure detection classification model, in the following ways:

[0084] (1) Preprocess the obtained historical dataset of heart sound signals.

[0085] (2) Extract features from the preprocessed historical dataset of heart sounds to generate a two-dimensional feature history array set of Mel frequency cepstral coefficients.

[0086] (3) Generate a first heart sound time-domain two-dimensional feature history set and a second heart sound time-domain two-dimensional feature history set based on the Mel frequency cepstral coefficient two-dimensional feature history set; the Mel frequency cepstral coefficient two-dimensional feature history set, the first heart sound time-domain two-dimensional feature history set and the second heart sound time-domain two-dimensional feature history set constitute a heart sound signal training dataset.

[0087] (4) Use the heart sound signal training dataset to train a multi-input temporal convolutional neural network model to construct a blood pressure detection classification model.

[0088] In this embodiment, the dataset used is a labeled heart sound signal training dataset. The label represents the blood pressure status corresponding to the heart sound signal data, including three conditions: normal blood pressure, high blood pressure, and low blood pressure. Normal blood pressure is defined as systolic blood pressure between 90-139 mmHg and diastolic blood pressure between 60-89 mmHg; high blood pressure is defined as systolic blood pressure >139 mmHg or diastolic blood pressure >89 mmHg; low blood pressure is defined as systolic blood pressure <90 mmHg or diastolic blood pressure <60 mmHg. Before model training, the heart sound signal training dataset is divided into a training set, a validation set, and a test set. The training set is used for model training, the validation set is used for testing the model's performance during training, and the test set is used for testing the model's performance after training. When training the multi-input temporal convolutional neural network model (MTCNN), the Adam optimizer is used, the total number of training epochs is set to 300, and the batch size is set to 64. At the end of each training cycle, the Multi-Input Temporal Convolutional Neural Network (MTCNN) model is tested on a validation set. When the validation accuracy improves, the model parameters are saved. Furthermore, the initial learning rate for MTCNN is 0.001. If the best accuracy on the validation set does not improve over 25 consecutive training cycles, the learning rate is reduced by a factor of 0.1, down to a minimum of 1e-8. The loss function used during MTCNN training is the cross-entropy loss function, most commonly used in classification tasks. Finally, the model training is complete, resulting in a trained MTCNN model used to construct a blood pressure detection classification model, serving as a blood pressure classifier.

[0089] In this embodiment, as Figure 4 The diagram illustrates the structure of a multi-input temporal convolutional neural network (TCNN) model in an embodiment of this application. The TCNN model includes three TCNNs, a concatenation layer, and a fully connected layer (FC). The TCNNs are merged through the concatenation layer, and the concatenation layer concatenates the one-dimensional feature arrays output by each TCNN before outputting them to the FC layer for computation. The three TCNNs are TCNN_1, TCNN_2, and TCNN_3. TCNN_1 inputs a one-dimensional temporal feature array of multi-channel Mel-frequency cepstral coefficients (MFCCs), TCNN_2 inputs a one-dimensional temporal feature array of multi-channel first heart sound (S1), and TCNN_3 inputs a one-dimensional temporal feature array of multi-channel second heart sound (S2).

[0090] In this embodiment, as Figure 5 The diagram illustrates the structure of a temporal convolutional neural network in an embodiment of this application. The temporal convolutional neural network includes: a first convolutional layer, a first normalized activation layer (BN layer + ReLU6), a temporal convolutional layer, and a global average pooling layer (Global_Avg_PooL). Each temporal convolutional layer includes multiple temporal convolutional blocks (TC-Blocks). Each TC-Block is composed of a second convolutional layer and a second normalized activation layer (BN layer + ReLU6). Both the first and second convolutional layers are one-dimensional convolutional layers. The first convolutional layer is a one-dimensional convolutional layer (Conv) with a kernel size of 1×3, and the second convolutional layer is a one-dimensional convolutional layer (Conv) with a kernel size of 1×9. The temporal convolutional neural network TCNN_1 has 16 temporal convolutional blocks, TCNN_2 has 8 temporal convolutional blocks, and TCNN_3 has 8 temporal convolutional blocks.

[0091] In this embodiment, three types of features—a one-dimensional time-domain feature array of multi-channel Mel-frequency cepstral coefficients, a one-dimensional time-domain feature array of multi-channel first heart sound, and a one-dimensional time-domain feature array of multi-channel second heart sound—are input into the corresponding TCNN_1, TCNN_2, and TCNN_3, respectively. First, a 1×3 first convolutional layer (Conv) is used to perform a convolution operation on the input feature array, thereby further compressing the input feature array. Then, the compressed features are input into the first normalized activation layer, where the batch normalization (BN) layer performs batch normalization. This BN layer operation can reduce overfitting to a certain extent and improve training speed. Next, the output of the BN layer is input into the ReLU6 activation function for non-linear operation. Since convolution is essentially a linear operation in the form of multiplication and accumulation, it cannot perform non-linear mapping, resulting in insufficient representation and fitting ability. Therefore, an additional activation function is needed to introduce non-linearity into the convolutional neural network, thereby enhancing the model's representation and fitting ability. The output of the ReLU6 activation function is then fed into the first temporal convolutional block (TC-Block), where a 1×9 convolution operation is performed, along with calculations of the BN layer and ReLU6 activation function in the second normalized activation layer. Subsequent temporal convolutional blocks (TC-Blocks) are calculated sequentially. Each TC-Block has two stride types: 1 and 2, representing the stride set for the convolutional layers within that block. Finally, the output of the last TC-Block is obtained, which is a multi-channel one-dimensional temporal feature. This multi-channel one-dimensional temporal feature is a two-dimensional feature, which is then global average pooled (Global_Avg_PooL) to reduce its dimensionality to a one-dimensional feature array. This completes the calculations for each TCNN, resulting in three one-dimensional feature arrays, which are then concatenated into a single one-dimensional vector using a vector concatenation layer. Finally, after computation by the fully connected layer (FC), the final blood pressure classification result is obtained. The model outputs three probability values, i.e., a 1×3 vector, corresponding to the labels of normal blood pressure, high blood pressure, and low blood pressure, respectively. The label corresponding to the maximum value is taken as the final blood pressure classification result and displayed. For example, if the maximum value is the first value of the 1×3 vector, the output is normal blood pressure; if the second value is the largest, the output is high blood pressure; and if the third value is the largest, the output is low blood pressure.

[0092] It is worth noting that the blood pressure detection classification system based on neural networks in this application is a portable blood pressure detection system based on a multi-input temporal convolutional neural network model. Compared with traditional blood pressure monitoring methods, it does not require going to a professional measurement institution or purchasing blood pressure measurement equipment. It has low requirements for the professionalism of the equipment and the data collector. It only requires a mobile phone microphone to collect real-time heart sound signal data and perform blood pressure detection. The data collection process is simple, the data collection difficulty is low, and it is more convenient and faster.

[0093] Furthermore, this application performs Mel-frequency cepstral coefficient (MFCC) feature extraction, first heart sound time-domain feature extraction, and second heart sound time-domain feature extraction on the collected real-time heart sound signal data. The extracted MFCC feature data, first heart sound (S1) time-domain feature data, and second heart sound (S2) time-domain feature data are used as model feature inputs, encompassing both time-domain and frequency-domain features. Furthermore, all input features have a certain mapping relationship with systolic blood pressure, diastolic blood pressure, and mean blood pressure. Compared to traditional methods that often use only one time-domain or frequency-domain feature, this system uses more comprehensive feature inputs, simultaneously capturing the frequency components of the heart sound signal and the specific time points of cardiac activity, achieving comprehensive coverage of time-frequency domain information. By using a multi-input time-domain convolutional neural network model to process and fuse these three feature inputs, the system can accurately detect the user's current blood pressure status, offering convenience, speed, and accurate results.

[0094] Furthermore, this application implements an algorithm for extracting the frame containing the peak heart sound point based on MFCC_0 (i.e., steps S301-S304). The algorithm is based on the first row feature vector (MFCC_0) in the two-dimensional feature array of Mel-frequency cepstral coefficients (MFCC), where MFCC_0 represents the total energy of the audio frame. According to a preset high energy threshold and a preset low energy threshold, the algorithm calculates the number of first frames in which each peak point satisfies the preset high energy threshold within a preset range, and the number of second frames in which each low peak point satisfies the preset low energy threshold within a preset range. The algorithm then determines the sum of the first frame counts and the sum of the second frame counts. Based on the determination result, the algorithm can accurately identify the signal frame containing the peak point of the first heart sound (S1) and the signal frame containing the peak point of the second heart sound (S2) in the heart sound signal, thereby generating a two-dimensional time-domain feature array of the first heart sound and a two-dimensional time-domain feature array of the second heart sound, which are then used as feature inputs to the model. Compared with traditional heart sound peak extraction algorithms, the heart sound peak point extraction algorithm of this application does not need to find a specific point, but only needs to find the signal frame where the peak point is located. It can effectively avoid the influence of noise, is simpler and more accurate, has stronger fault tolerance, and can accurately distinguish between the first heart sound (S1) and the second heart sound (S2).

[0095] Finally, this application constructs a blood pressure detection classification model by training a multi-input temporal convolutional neural network (MTCNN) model. This model possesses stronger learning and generalization capabilities, enabling more accurate blood pressure classification results. It is applicable to any population and does not require pre-training based on user data for accurate and rapid blood pressure detection. Furthermore, compared to traditional neural network models, the proposed MTCNN model simultaneously uses Mel-frequency cepstral coefficients (MFCC) feature data, first heart sound (S1) temporal feature data, and second heart sound (S2) temporal feature data as model inputs. This allows for a more comprehensive learning of blood pressure-related features in the heart sound signal, resulting in stronger robustness and generalization ability. In addition, the MTCNN model uses a large number of one-dimensional temporal convolutions instead of traditional two-dimensional convolutions, paying more attention to temporal information and thus better learning the temporal features of the heart sound signal. It also compresses frequency domain information, reducing computational load. Ultimately, the multi-input temporal convolutional neural network model (MTCNN) proposed in this application can efficiently and accurately complete the blood pressure classification task with extremely low parameter and computational requirements.

[0096] In the embodiments of this application, terms such as "first" and "second" are used to distinguish identical or similar items with essentially the same function and effect. For example, the first frame number and the second frame number are only used to distinguish different frame numbers and do not limit their order. Those skilled in the art will understand that terms such as "first" and "second" do not limit the quantity or execution order, and that terms such as "first" and "second" do not necessarily imply that they are different.

[0097] It should be noted that, in the embodiments of this application, the words "exemplary" or "for example" indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of words such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0098] In this application embodiment, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be single or multiple.

[0099] It should be understood that the specific process of each module performing the above-mentioned steps has been described in detail in the above method embodiments, and will not be repeated here for the sake of brevity.

[0100] It should also be understood that the module division in the embodiments of this application is illustrative and only represents a logical functional division; in actual implementation, there may be other division methods. Furthermore, the functional modules in the various embodiments of this application can be integrated into a single processor, exist as separate physical entities, or be integrated into a single module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0101] like Figure 6 The diagram illustrates a flowchart of the neural network-based blood pressure detection and classification method in an embodiment of this application. The neural network-based blood pressure detection and classification method includes the following steps:

[0102] S601: Preprocess the acquired real-time heart sound signal data.

[0103] S602: Extract features from the preprocessed real-time heart sound signal data to generate a two-dimensional feature array of Mel frequency cepstral coefficients.

[0104] S603: Generate a first two-dimensional feature array of heart sounds in the time domain and a second two-dimensional feature array of heart sounds in the time domain based on the two-dimensional feature array of Mel frequency cepstral coefficients.

[0105] S604: Based on the two-dimensional feature array of Mel frequency cepstral coefficients, the two-dimensional feature array of the first heart sound in the time domain, and the two-dimensional feature array of the second heart sound in the time domain, and based on the pre-built blood pressure detection classification model, generate a blood pressure detection classification result and output the classification result.

[0106] Figure 7 This is a schematic block diagram of the electronic terminal provided in an embodiment of this application. Figure 7As shown, the electronic terminal 700 includes at least one processor 701, a memory 702, at least one network interface 703, and a user interface 705. The various components in the device are coupled together via a bus system 704. It is understood that the bus system 704 is used to implement communication between these components. In addition to a data bus, the bus system 704 also includes a power bus, a control bus, and a status signal bus. However, for clarity, in… Figure 7 The general will label all buses as bus systems.

[0107] The user interface 705 may include a monitor, keyboard, mouse, trackball, clicker, button, touchpad, or touch screen.

[0108] It is understood that memory 702 can be volatile memory or non-volatile memory, or both. Non-volatile memory can be read-only memory (ROM) or programmable read-only memory (PROM), used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM) and synchronous static random access memory (SSRAM). The memories described in the embodiments of this invention are intended to include, but are not limited to, these and any other suitable categories of memory.

[0109] In this embodiment of the invention, the memory 702 is used to store various categories of data to support the operation of the electronic terminal 700. Examples of this data include: any executable program for operation on the electronic terminal 700, such as the operating system 7021 and application program 7022; the operating system 7021 contains various system programs, such as the framework layer, core library layer, driver layer, etc., for implementing various basic services and handling hardware-based tasks. The application program 7022 may contain various applications, such as a media player, browser, etc., for implementing various application services. The implementation of the neural network-based blood pressure detection and classification method provided in this embodiment of the invention can be included in the application program 7022.

[0110] The methods disclosed in the above embodiments of the present invention can be applied to or implemented by processor 701. Processor 701 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the integrated logic circuit of the hardware in processor 701 or by instructions in software form. The processor 701 may be a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Processor 701 can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of the present invention. General-purpose processor 701 may be a microprocessor or any conventional processor, etc. The steps of the accessory optimization method provided in the embodiments of the present invention can be directly reflected as being executed by a hardware decoding processor, or being executed by a combination of hardware and software modules in the decoding processor. The software module may be located in a storage medium, which is located in memory. The processor reads the information in the memory and combines it with its hardware to complete the steps of the aforementioned method.

[0111] In an exemplary embodiment, the electronic terminal 700 may be used by one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), or complex programmable logic devices (CPLDs) to execute the aforementioned method.

[0112] According to the method provided in the embodiments of this application, this application also provides a computer program product, which includes: computer program code, which, when run on a computer, causes the computer to execute... Figure 6 The method of any of the embodiments shown.

[0113] According to the method provided in the embodiments of this application, this application also provides a computer-readable storage medium storing program code, which, when executed on a computer, causes the computer to perform... Figure 6 The method of any of the embodiments shown.

[0114] As used in this specification, the terms "component," "module," "system," etc., are used to refer to computer-related entities, hardware, firmware, combinations of hardware and software, software, or software in execution. For example, a component can be, but is not limited to, a process running on a processor, a processor, an object, an executable file, an execution thread, a program, and / or a computer. As illustrated, applications running on computing devices and computing devices can both be components. One or more components may reside in a process and / or an execution thread, and components may be located on a single computer and / or distributed among two or more computers. Furthermore, these components can be executed from various computer-readable media on which various data structures are stored. Components can communicate, for example, via local and / or remote processes based on signals having one or more data packets (e.g., data from two components interacting with another component between a local system, a distributed system, and / or a network, such as the Internet interacting with other systems via signals).

[0115] Those skilled in the art will recognize that the various illustrative logical blocks and steps described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0116] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0117] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0118] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0119] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0120] In the above embodiments, the functions of each functional unit can be implemented entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. A computer program product includes one or more computer instructions (programs). When the computer program instructions (programs) are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., high-density digital video discs, DVDs), or semiconductor media (e.g., solid-state disks, SSDs, etc.).

[0121] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0122] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0123] In summary, this application provides a blood pressure detection and classification system, method, medium, and terminal based on a neural network. It extracts Mel-frequency cepstral coefficient features, first heart sound time-domain features, and second heart sound time-domain features from the acquired real-time heart sound signal data. The extracted Mel-frequency cepstral coefficient features, first heart sound time-domain features, and second heart sound time-domain features are used as model feature inputs, encompassing both time-domain and frequency-domain features, achieving comprehensive coverage of time-frequency domain information. By using a multi-input time-domain convolutional neural network model to process and fuse these three feature inputs, the system can accurately detect the user's current blood pressure status, offering convenience, speed, and accurate results. This invention implements a heart sound peak point extraction algorithm based on MFCC_0, which effectively avoids the influence of noise, is simpler and more accurate, has stronger fault tolerance, and can accurately distinguish between the first and second heart sounds. This application constructs a blood pressure detection classification model by training a multi-input temporal convolutional neural network model. This model possesses significantly enhanced learning and generalization capabilities, enabling efficient and accurate blood pressure classification with minimal parameters and computational load. It yields more accurate blood pressure classification results and is applicable to any population group. Furthermore, it eliminates the need for pre-training based on user data, allowing for accurate and rapid blood pressure detection. Therefore, this application effectively overcomes the shortcomings of existing technologies and possesses high industrial applicability.

[0124] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this application should still be covered by the claims of this application.

Claims

1. A blood pressure detection and classification system based on neural network, characterized in that: include: A preprocessing module, used for preprocessing the acquired real-time data of heart sound signals; The first feature extraction module is used to extract features from the preprocessed real-time heart sound signal data and generate a two-dimensional feature array of Mel-frequency cepstral coefficients; A second feature extraction module, used for generating a first heart sound time domain two-dimensional feature array and a second heart sound time domain two-dimensional feature array according to the Mel-frequency cepstral coefficient two-dimensional feature array; The blood pressure detection classification module is used to generate a blood pressure detection classification result and output the classification result based on the Mel-frequency cepstral coefficient two-dimensional feature array, the first heart sound time domain two-dimensional feature array and the second heart sound time domain two-dimensional feature array, and based on a pre-built blood pressure detection classification model.

2. The blood pressure detection and classification system based on neural network according to claim 1, characterized in that: The second feature extraction module generates the first heart sound time domain two-dimensional feature array and the second heart sound time domain two-dimensional feature array according to the Mel-frequency cepstral coefficient two-dimensional feature array, including: According to the preset window length, traverse the first row of feature vectors in the two-dimensional feature array of Mel-frequency cepstral coefficients in turn to filter out all the maximum points in the first row of feature vectors, and record the signal frames where each of the maximum points is located and the corresponding energy values; According to the preset high energy threshold, the preset low energy threshold and the energy value corresponding to each of the maximum value points, all high peak points and all low peak points in the first row of feature vectors are screened out, and the signal frames where each of the high peak points is located and the corresponding energy value, and the signal frames where each of the low peak points is located and the corresponding energy value are recorded; Calculating a first number of frames in which each of the high peak points satisfies a preset high energy threshold within a preset value range, and calculating a second number of frames in which each of the low peak points satisfies a preset low energy threshold within a preset value range; The magnitude of the sum of the first frame number and the sum of the second frame number is determined, and based on the determination result, a first heart sound time-domain two-dimensional feature array and a second heart sound time-domain two-dimensional feature array are generated.

3. The blood pressure detection and classification system based on neural network according to claim 1, characterized in that: The blood pressure detection classification module generates a blood pressure detection classification result according to the Mel frequency cepstral coefficient two-dimensional feature array, the first heart sound time domain two-dimensional feature array and the second heart sound time domain two-dimensional feature array, and based on a pre-constructed blood pressure detection classification model, and outputs the classification result in the following manner: Performing dimension-changing processing on the two-dimensional feature array of Mel-frequency cepstral coefficients, the two-dimensional feature array of the first heart sound in the time domain, and the two-dimensional feature array of the second heart sound in the time domain, respectively, to generate a multi-channel one-dimensional time-domain feature array of Mel-frequency cepstral coefficients, a multi-channel one-dimensional time-domain feature array of the first heart sound, and a multi-channel one-dimensional time-domain feature array of the second heart sound; The multi-channel Mel-frequency cepstral coefficient one-dimensional time domain feature array, the multi-channel first heart sound one-dimensional time domain feature array and the multi-channel second heart sound one-dimensional time domain feature array are respectively input into a pre-constructed blood pressure detection classification model to generate a blood pressure detection classification result and output the classification result.

4. The blood pressure detection and classification system based on neural network according to claim 1, characterized in that: It also includes a model building module, which is used to build a blood pressure detection classification model in the following ways: Preprocessing the acquired heart sound signal historical data set; Perform feature extraction on the preprocessed heart sound signal history data set to generate a two-dimensional feature history array set of Mel frequency cepstral coefficients; Generate a first heart sound time domain two-dimensional feature history array set and a second heart sound time domain two-dimensional feature history array set according to the Mel-frequency cepstral coefficient two-dimensional feature history array set; the Mel-frequency cepstral coefficient two-dimensional feature history array set, the first heart sound time domain two-dimensional feature history array set and the second heart sound time domain two-dimensional feature history array set constitute a heart sound signal training data set; The heart sound signal training data set is used to train a multi-input time-domain convolutional neural network model to build a blood pressure detection classification model.

5. The blood pressure detection and classification system based on neural network according to claim 4, characterized in that: The multi-input time-domain convolutional neural network model includes three time-domain convolutional neural networks, a vector splicing layer and a fully connected layer; each time-domain convolutional neural network is merged through the vector splicing layer, and the one-dimensional feature arrays output by each time-domain convolutional neural network are spliced ​​through the vector splicing layer and output to the fully connected layer for calculation.

6. The blood pressure detection and classification system based on neural network according to claim 5, characterized in that: The time domain convolutional neural network includes: a first convolutional layer, a first normalized activation layer, a time domain convolutional layer and a global average pooling layer; the time domain convolutional layer includes a plurality of time domain convolutional blocks; the time domain convolutional block includes a second convolutional layer and a second normalized activation layer; The first convolution layer and the second convolution layer are both one-dimensional convolution layers.

7. The blood pressure detection and classification system based on neural network according to claim 1, characterized in that: The first feature extraction module extracts features from the preprocessed real-time heart sound signal data to generate a two-dimensional feature array of Mel-frequency cepstrum coefficients, including: According to a preset frame length and a preset frame shift, the pre-processed real-time heart sound signal data is divided into a plurality of signal frames; According to the preset window function, weighting operations are performed on each segmented signal frame; Performing fast Fourier transform processing on each signal frame after the weighting operation is completed to obtain spectrum data corresponding to each signal frame; Performing filtering on the spectrum data corresponding to each signal frame to obtain the filtering coefficient corresponding to each signal frame; Logarithm processing is performed on the filter coefficients corresponding to each signal frame respectively, and discrete cosine transform processing is performed on the filter coefficients corresponding to each signal frame after logarithm processing to generate a two-dimensional feature array of Mel-frequency cepstral coefficients.

8. A blood pressure detection and classification method based on a neural network, characterized in that: include: Preprocess the acquired real-time data of heart sound signals; Perform feature extraction on the preprocessed real-time heart sound signal data to generate a two-dimensional feature array of Mel-frequency cepstral coefficients; Generate a first heart sound time domain two-dimensional feature array and a second heart sound time domain two-dimensional feature array according to the Mel-frequency cepstral coefficient two-dimensional feature array; According to the Mel-frequency cepstral coefficient two-dimensional feature array, the first heart sound time-domain two-dimensional feature array and the second heart sound time-domain two-dimensional feature array, and based on a pre-constructed blood pressure detection classification model, a blood pressure detection classification result is generated and the classification result is output.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the blood pressure detection and classification method based on a neural network as described in claim 8 is implemented.

10. An electronic terminal comprising a memory, a processor and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the blood pressure detection and classification method based on neural network as described in claim 8.

Images