Heart sound data processing method and system based on time-frequency feature map

Abstract

The application discloses a heart sound data processing method and system based on a time-frequency feature map, and relates to the technical field of electric digital data processing. The heart sound data processing method based on the time-frequency feature map comprises the following steps: heart sound stable division, optimized identification processing and noise identification verification. The application obtains a heart sound stable difference characteristic result by performing feature stable difference division on a time-frequency feature map obtained based on a heart sound signal, then performs heart sound optimization processing in combination with the heart sound stable difference characteristic result and obtains a noise level identification result, finally verifies the noise level identification result, if the verification is successful, the noise level identification result is fed back, otherwise, the noise level identification result is updated, so that the heart sound signal noise identification effect is more accurate, and the problem that the heart sound signal stability difference is not fully considered in the heart sound signal noise identification process based on the time-frequency feature map in the prior art is solved.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing technology, and in particular to a method and system for processing heart sound data based on time-frequency feature maps. Background Technology

[0002] With the rapid development of the healthcare industry, especially the rise of personalized medicine and remote health monitoring, the analysis of heart sound signals plays an increasingly important role in clinical diagnosis. Traditional methods of diagnosing heart disease, such as auscultation and electrocardiography, rely on the experience and judgment of doctors, which can easily lead to misdiagnosis or missed diagnosis due to subjectivity. With the advancement of artificial intelligence and signal processing technology, automated heart sound and murmur recognition has become an important way to improve diagnostic efficiency and accuracy. Heart sound signals are low-frequency biological signals, usually containing multiple components. Due to the diversity and complexity of heart sounds, accurately identifying murmurs and distinguishing between healthy and pathological signals becomes difficult. Time-frequency feature maps, by converting heart sound signals into time and frequency domain representations, show the intensity and changes of different frequency components. This representation method helps to observe various features in heart sounds, but it also faces some problems. For example, time-frequency feature maps are obtained through methods such as short-time Fourier transform, which presents a balance problem between time and frequency. Too long a window can lead to the loss of temporal information, while too short a window can affect frequency resolution, thus affecting the recognition of murmurs. In heart sound signal analysis, especially in the application of time-frequency feature maps, obtaining a fully labeled and high-quality heart sound dataset is a challenge. Insufficient labeled data can lead to inadequate model training, thus affecting recognition performance. While heart sound murmur identification using time-frequency feature maps faces technical and data challenges, the research and application prospects in this field are broad due to the continuous development of signal processing technology, deep learning algorithms, wearable devices, and telemedicine technologies. With the accumulation of larger-scale, high-quality datasets, future research and technologies will be able to overcome current difficulties, providing more accurate and efficient methods for heart sound murmur identification, thus promoting the early prevention and diagnosis of heart disease.

[0003] Existing time-frequency feature mapping techniques can provide important information about heart health by acquiring and analyzing heart sound signals. Murmur identification is a crucial step in heart sound analysis, aiming to identify abnormal sounds (such as heart palpitations, murmurs, or other unusual noises) within the heart sound signal. Heart sound signals are typically acquired using digital stethoscopes, wearable devices (such as chest patch devices, smartwatches, etc.), or electronic stethoscopes. After acquisition, heart sound signals usually require preprocessing to remove noise and interference, enhancing signal quality. To extract time-frequency features from heart sound signals, techniques such as short-time Fourier transform, wavelet transform, or Mel-frequency cepstral coefficients are commonly used to convert the heart sound signal into a time-frequency domain spectrogram. With the development of deep learning, convolutional neural networks, long short-term memory networks, and recurrent neural networks have gradually become mainstream methods for heart sound and murmur identification. Deep learning models can automatically learn features from the spectrogram, avoiding the complexity of manual feature selection in traditional methods, and possessing stronger expressive power and robustness.

[0004] For example, the invention patent announcement CN111860246B discloses a data augmentation method for heart sound signal classification for deep convolutional neural networks, which includes: preprocessing the original training heart sound data; converting the preprocessed one-dimensional heart sound signal into a two-dimensional MFSC feature map; multiplying the MFSC feature map by a mask function to randomly mask its frequency and time domains; expanding and balancing the number of samples in each category to obtain a new training dataset; and training a deep convolutional neural network with the new training dataset.

[0005] For example, the invention patent announcement CN112036467B discloses an abnormal heart sound recognition method and apparatus based on a multi-scale attention neural network, which includes: preprocessing the collected raw heart sound signals and using the preprocessed heart sound signals as training samples; annotating the heart sound quality of the training samples; training an abnormal heart sound recognition model based on the training samples and their annotations; inputting the heart sound data to be detected into the trained abnormal heart sound recognition model to obtain a heart sound quality prediction result; and identifying abnormal heart sounds based on the heart sound quality prediction result.

[0006] However, in the process of implementing the inventive technical solution in the embodiments of this application, it was found that the above-mentioned technology has at least the following technical problems:

[0007] Time-frequency feature maps are used to extract and analyze various features in heart sound signals. However, when using time-frequency feature maps (such as short-time Fourier transform) to process heart sound data, the non-stationarity of heart sound signals in a short period of time is often ignored. In particular, the frequency and amplitude of heart sound signals may change significantly in different cardiac cycles (such as different stages of atrial and ventricular contraction), which is insufficient to accurately reflect the actual characteristics of heart sound signals.

[0008] In existing technologies, in actual heart sound signals, noise (such as background noise, breathing sounds, muscle vibrations, etc.) and heart sound signals often overlap in the frequency spectrum. Due to the non-stationarity of heart sound signals, their spectral characteristics change over time, making it more difficult to distinguish between heart sound signals and noise signals. If the non-stationarity of heart sound signals is not fully considered, the time-frequency feature map may mistake noise for part of the heart sound or fail to effectively separate the two, leading to a decrease in recognition accuracy. This indicates a problem with the time-frequency feature map-based heart sound noise recognition process not fully considering the stationarity differences of heart sound signals. Summary of the Invention

[0009] This invention provides a method and system for processing heart sound data based on time-frequency feature maps, which solves the problem in the prior art that the differences in the stability of heart sound signals are not fully considered in the process of identifying heart sound noise based on time-frequency feature maps, and achieves more accurate identification of heart sound noise.

[0010] This invention provides a method for processing heart sound data based on time-frequency feature maps, comprising the following steps: S1, acquiring heart sound signals within a preset time interval, and performing feature stationary difference segmentation on the time-frequency feature map obtained based on the heart sound signals to obtain heart sound stationary difference characteristic results; S2, performing heart sound optimization processing based on the heart sound stationary difference characteristic results to obtain noise level identification results, wherein the heart sound optimization processing means optimizing the heart sound signal input in the noise identification process based on the heart sound stationary difference characteristic results to improve the accuracy of heart sound signal noise identification; S3, verifying the noise level identification results, and if the verification is successful, feeding back the noise level identification results; otherwise, updating the noise level identification results.

[0011] Furthermore, the time-frequency feature map obtained based on the heart sound signal is subjected to feature stationary difference segmentation. The specific steps are as follows: Feature segmentation data of the heart sound signal within a preset time interval is obtained. The feature segmentation data includes signal autocorrelation value, signal stationarity test root, signal frequency value, signal average frequency, and average power spectral density. After performing difference averaging approximation operation on the signal autocorrelation value of the heart sound signal at the preset acquisition monitoring time and the left adjacent acquisition monitoring time, the signal stationarity test root approximation operation result is subjected to time-domain stationarity interaction processing to obtain the time-domain stationarity. The time-domain stationarity interaction processing is used to describe the interaction between the signal autocorrelation value difference averaging approximation operation result and the signal stationarity test root approximation operation result. After performing difference analysis between the signal frequency value and the signal average frequency, the average power spectral density approximation calculation result is subjected to frequency domain stationarity interaction processing to obtain the frequency domain stationarity. The frequency domain stationarity interaction processing is used to describe the interaction between the signal frequency value difference analysis result and the average power spectral density approximation calculation result. The time domain stationarity, frequency domain stationarity, and corresponding stationarity compensation are weighted and coupled to obtain the heart sound stationarity partition value. The heart sound stationarity partition value is used to quantitatively evaluate the stationarity of the center sound signal feature in the time-frequency feature map. The stationarity compensation includes time domain stationarity compensation and frequency domain stationarity compensation. The heart sound stationarity partition value represents the quantitative data of the time domain stationarity and frequency domain stationarity jointly evaluating the stationarity of the center sound signal feature in the time-frequency feature map.

[0012] Furthermore, the heart sound stability difference characteristic results are obtained, and the specific process is as follows: Based on the preset heart sound stability characteristic threshold range obtained from the preset database, it is determined whether the obtained heart sound stability segmentation value is within the preset heart sound stability characteristic threshold range; if the heart sound stability segmentation value is within the preset heart sound stability characteristic threshold range, the heart sound stability difference characteristic result is recorded as a stable heart sound signal; if the heart sound stability segmentation value is not within the preset heart sound stability characteristic threshold range, the heart sound stability difference characteristic result is recorded as a non-stationary heart sound signal; the heart sound stability difference characteristic result includes stable heart sound signals and non-stationary heart sound signals.

[0013] Furthermore, the specific steps for optimizing heart sounds based on the results of the heart sound stability difference characteristics are as follows: If the heart sound stability difference characteristic result is a stable heart sound signal, then the stable heart sound signal is subjected to stable heart sound optimization processing; stable heart sound optimization processing means optimizing the input of a stable heart sound signal based on the heart sound stability difference characteristic result to improve the accuracy of the heart sound signal in the noise recognition process; if the heart sound stability difference characteristic result is a non-stationary heart sound signal, then the non-stationary heart sound signal is subjected to non-stationary heart sound optimization processing; non-stationary heart sound optimization processing means optimizing the input of a non-stationary heart sound signal based on the heart sound stability difference characteristic result to improve the accuracy of the heart sound signal in the noise recognition process; heart sound optimization processing includes stable heart sound optimization processing and non-stationary heart sound optimization processing.

[0014] Further, the specific process for optimizing the stable heart sound signal is as follows: the stable heart sound signal is divided into short-time stable heart sound signals with equal lengths of the stable time window. Matching detection is performed based on the overlap between the short-time stable heart sound signals and the reference known noise signal to obtain an overlap matching detection value. This value is used to quantify the degree of overlap between the short-time stable heart sound signal and the reference known noise signal. The stable time window length represents the result of mapping the deviation between the real-time stable heart sound segmentation value and the stable heart sound reference threshold to a stable segmentation length mapping set in a preset database. This mapping set represents the mapping relationship between the stable heart sound segmentation value, the deviation from the stable heart sound reference threshold, and the stable time window length. If the deviation between the overlap matching detection value and the reference overlap matching threshold is within a preset allowable deviation range obtained from the preset database, the noise level recognition input priority of the short-time stable heart sound signal is obtained by sorting the input confidence scores of the short-time stable heart sound signals in descending order. A preset number of short-time stable heart sound signals are then sorted according to this noise level recognition input priority. A short-term stable heart sound signal is input into a pre-constructed murmur level recognition model to obtain the murmur level recognition result. The input confidence score represents the result of mapping the deviation between the real-time overlap match detection value and the reference overlap match threshold to an input confidence mapping set in a preset database. The input confidence mapping set represents the mapping relationship between the deviation between the overlap match detection value and the reference overlap match threshold and the input confidence score. If the deviation between the overlap match detection value and the reference overlap match threshold is not within the preset allowable deviation range obtained from the preset database, principal component analysis is performed based on the feature dimensionality reduction strength of the short-term stable heart sound signal. A preset number of short-term stable heart sound signals after principal component analysis dimensionality reduction are input into the pre-constructed murmur level recognition model to obtain the murmur level recognition result. The feature dimensionality reduction strength represents the result of mapping the deviation between the real-time overlap match detection value and the reference overlap match threshold to a dimensionality reduction strength mapping set in a preset database. The dimensionality reduction strength mapping set represents the mapping relationship between the deviation between the overlap match detection value and the reference overlap match threshold and the feature dimensionality reduction strength.

[0015] Furthermore, based on the overlap between the short-time stationary heart sound signal and the reference known noise signal, a matching detection is performed to obtain the overlap matching detection value. The specific steps are as follows: The overlap detection quantization data between the short-time stationary heart sound signal and the reference known noise signal is obtained. The overlap detection quantization data includes the detection cross-correlation coefficient, the detection similarity coefficient, and the spectral overlap coefficient. The detection cross-correlation coefficient is used to describe the similarity between the short-time stationary heart sound signal and the reference known noise signal in the time domain; the detection similarity coefficient is used to describe the structural similarity between the short-time stationary heart sound signal and the reference known noise signal; and the spectral overlap coefficient is used to describe the similarity between the short-time stationary heart sound signal and the reference known noise signal in the frequency domain. The overlap detection quantization data and the corresponding detection compensation amount are weighted and coupled to obtain the overlap matching detection value. The detection compensation amount includes the cross-correlation detection compensation amount, the similarity detection compensation amount, and the spectral overlap compensation amount. The overlap matching detection value represents the quantization data of the overlap detection quantization data for the overlap matching detection of the short-time stationary heart sound signal and the reference known noise signal.

[0016] Furthermore, the specific process for optimizing non-stationary heart sound signals is as follows: First, determine whether the deviation between the stationary heart sound segmentation value and the stationary heart sound reference threshold is greater than a stationary deviation setting value obtained from a preset database. If the deviation is greater, input the deviation into a heart sound sampling mapping set in the preset database to obtain the heart sound signal sampling frequency. The heart sound signal is then acquired at this sampling frequency. The heart sound sampling mapping set represents the mapping relationship between the deviation of the stationary heart sound segmentation value and the stationary deviation setting value and the heart sound signal sampling frequency. Second, if the deviation is not greater than the stationary deviation setting value obtained from the preset database, divide the non-stationary heart sound signal into short-time non-stationary heart sound signals with equal non-stationary time window lengths. The non-stationary time window length represents the length of the non-stationary segmentation length mapping set in the preset database. The results of the analysis show that the non-stationary partition length mapping set represents the mapping relationship between the deviation of the stationary heart sound partition value and the stationary heart sound reference threshold and the length of the non-stationary time window. The signal energy value of the short-term non-stationary heart sound signal is obtained, and it is determined whether the signal energy value of the short-term non-stationary heart sound signal is greater than the reference signal energy threshold obtained from the preset database. If the signal energy value of the short-term non-stationary heart sound signal is greater than the reference signal energy threshold, the short-term non-stationary heart sound signal is marked as a valid heart sound signal; otherwise, it is marked as a waiting heart sound signal. It is determined whether the short-term non-stationary heart sound signals in the adjacent non-stationary time windows of the valid heart sound signal are all valid heart sound signals. The length of the adjacent non-stationary time window includes the left and right adjacent non-stationary time windows of the valid heart sound signal. If the short-term non-stationary heart sound signals in the adjacent non-stationary time window lengths of the valid heart sound signal are all valid heart sound signals, the valid heart sound signal is marked as a valid priority heart sound signal; otherwise, no marking is performed. The valid priority heart sound signal is input into the constructed noise level recognition model to obtain the noise level recognition result.

[0017] Furthermore, the noise level identification results are verified, and the specific process is as follows: If the heart sound stability difference characteristic result is a stable heart sound signal, then a preset multiple of short-time stable heart sound signals are input into the constructed noise level identification model to obtain the verified noise level identification result; if the heart sound stability difference characteristic result is a non-stationary heart sound signal, then a preset multiple of valid heart sound signals are input into the constructed noise level identification model to obtain the verified noise level identification result; it is then determined whether the conformity between the noise level identification result and the verified noise level identification result is greater than a preset conformity threshold obtained from a preset database; if the conformity between the noise level identification result and the verified noise level identification result is greater than the preset conformity threshold obtained from a preset database, then the verification is successful; otherwise, the noise level identification result is updated.

[0018] Furthermore, the specific steps for updating the noise level identification result are as follows: verify the noise level identification result. If the verification is still unsuccessful, verify the output result in sequence. If the verification is successful, update the noise level identification result to the output result of the verification. If the verification is still unsuccessful when the number of input heart sound signals reaches the preset maximum number, update the noise level identification result to identification abnormal.

[0019] This invention provides a heart sound data processing system based on time-frequency feature maps, comprising: a heart sound stationary segmentation module, an optimized recognition processing module, and a noise recognition verification module; the heart sound stationary segmentation module is used to acquire heart sound signals within a preset time interval, and perform feature stationary difference segmentation on the time-frequency feature map obtained based on the heart sound signals to obtain heart sound stationary difference characteristic results; the optimized recognition processing module is used to perform heart sound optimization processing based on the heart sound stationary difference characteristic results to obtain noise level recognition results, whereby the heart sound optimization processing means optimizing the heart sound signal input in the noise recognition process based on the heart sound stationary difference characteristic results to improve the accuracy of heart sound signal noise recognition; the noise recognition verification module is used to verify the noise level recognition results, and if the verification is successful, the noise level recognition results are fed back; otherwise, the noise level recognition results are updated.

[0020] One or more technical solutions provided in this invention have at least the following technical effects or advantages:

[0021] 1. By dividing the time-frequency feature map obtained from the heart sound signal into stationary differences, the stationary difference characteristics of the heart sound are obtained. Then, the heart sound optimization processing is performed based on the stationary difference characteristics of the heart sound, and the noise level identification result is obtained. Finally, the noise level identification result is verified. If the verification is successful, the noise level identification result is fed back; otherwise, the noise level identification result is updated. This realizes the analysis of the stationary difference characteristics of the heart sound signal and the optimization of the heart sound signal input in the noise identification process, thereby achieving more accurate noise identification of the heart sound signal. It effectively solves the problem in the existing technology that the stationary difference of the heart sound signal is not fully considered in the process of noise identification of the heart sound signal based on the time-frequency feature map.

[0022] 2. By combining the results of the heart sound stability difference characteristics, heart sound optimization processing is performed. If the heart sound stability difference characteristics result is a stable heart sound signal, then the stable heart sound signal is optimized for stability. If the heart sound stability difference characteristics result is a non-stable heart sound signal, then the non-stable heart sound signal is optimized for non-stable heart sound. This achieves the optimization of the difference in the input of the heart sound signal for noise recognition by combining the stability difference of the heart sound signal, thereby improving the input accuracy in the noise recognition process of the heart sound signal.

[0023] 3. By verifying the noise level identification result, if the verification is successful, the noise level identification result is updated to the output result of the verification; otherwise, the output result is verified sequentially. If the verification is still unsuccessful when the number of input heart sound signals reaches the preset maximum number, the noise level identification result is updated to identification abnormal. This achieves further verification of the noise level identification result, thereby improving the reliability of obtaining the noise level identification result from the heart sound signal noise identification. Attached Figure Description

[0024] Figure 1 A flowchart of a heart sound data processing method based on time-frequency feature maps provided in this application embodiment;

[0025] Figure 2 A step logic diagram of the heart sound data processing method based on time-frequency feature maps provided in the embodiments of this application;

[0026] Figure 3 This is a schematic diagram of the structure of the heart sound data processing system based on time-frequency feature maps provided in an embodiment of this application. Detailed Implementation

[0027] This application provides a method and system for processing heart sound data based on time-frequency feature maps. This solves the problem in existing technologies where the stationarity differences of heart sound signals are not fully considered during noise identification based on time-frequency feature maps. The method acquires heart sound signals within a preset time interval, then divides the time-frequency feature map obtained from the heart sound signals into stationary differences to obtain heart sound stationarity difference characteristic results. Next, heart sound optimization processing is performed based on the heart sound stationarity difference characteristic results to obtain noise level identification results. Finally, the noise level identification results are verified. If the verification is successful, the noise level identification results are fed back; otherwise, the noise level identification results are updated, thus achieving more accurate noise identification of heart sound signals.

[0028] The technical solution in this application embodiment addresses the problem that the differences in the stationarity of heart sound signals are not fully considered in the above-mentioned process of identifying heart sound murmurs based on time-frequency feature maps. The overall approach is as follows:

[0029] By dividing the time-frequency feature map obtained based on the heart sound signal into stationary differences, then optimizing the heart sound based on the stationary difference characteristics and obtaining the noise level identification result, and finally verifying the noise level identification result, a more accurate effect of identifying noise in the heart sound signal is achieved.

[0030] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0031] like Figure 1 The diagram shows a flowchart of a heart sound data processing method based on a time-frequency feature map provided in this application embodiment. The method includes the following steps: S1, acquiring heart sound signals within a preset time interval, and performing feature stationary difference division on the time-frequency feature map obtained based on the heart sound signals to obtain heart sound stationary difference characteristic results; S2, performing heart sound optimization processing based on the heart sound stationary difference characteristic results to obtain noise level recognition results. Heart sound optimization processing means optimizing the heart sound signal input in the noise recognition process based on the heart sound stationary difference characteristic results to improve the accuracy of heart sound signal noise recognition; S3, verifying the noise level recognition results. If the verification is successful, the noise level recognition results are fed back; otherwise, the noise level recognition results are updated.

[0032] In this embodiment, as Figure 2The diagram shows the step logic of the heart sound data processing method based on time-frequency feature maps provided in this application embodiment. Heart sound signals reflect the activity of the heart, and the heart activity itself has different physiological and pathological manifestations at different time points. For example, the working state of the heart during systole and diastole is different, and the characteristics of the heart sounds also change accordingly. In addition, pathological conditions such as heart disease and valvular heart disease can change the rhythm of the heart, blood flow, and vibration characteristics of heart tissue, resulting in abnormal waveforms of heart sounds, which in turn leads to non-stationarity of the signal. Furthermore, heart sounds not only include normal heart sounds (such as the first heart sound and the second heart sound), but may also contain murmurs. Murmurs usually appear at certain specific times and change with the changes in the condition. For example, murmurs produced by valvular heart disease often overlap with normal heart sounds, and their frequency and amplitude change with the heart's activity cycle, leading to non-stationarity of the signal. At the same time, murmurs may also come from environmental noise, equipment errors, patient movement or changes in body position, etc. Due to the inherent characteristics of the heart sound signal (such as the spectral overlap between heart sounds and murmurs), noise identification becomes more difficult.

[0033] In heart sound signal processing, the stationarity differences of heart sounds are crucial for accurate signal processing and murmur identification. This difference directly affects feature extraction, model training, and the final identification results; ignoring this difference may affect the accuracy of feature extraction, model training, and identification results. To address this issue, the algorithm in this application optimizes the heart sound signal input during murmur identification based on the feature analysis of heart sound signal stationarity differences. This effectively improves the precision of heart sound signal processing and the accuracy of murmur identification, thus providing a more reliable auxiliary diagnostic tool for clinical use.

[0034] Furthermore, the time-frequency feature map obtained based on the heart sound signal is divided into stationary differences. The specific steps are as follows:

[0035] A1. Obtain feature segmentation data of the inner sound signal at a preset time interval. The feature segmentation data includes signal autocorrelation value, signal stationarity test root, signal frequency value, signal average frequency, and average power spectral density. Specifically, the signal autocorrelation value is obtained using the autocorrelation function in the NumPy and SciPy libraries in Python; the signal stationarity test root is obtained using the statsmodels library in Python (adfuller function for ADF test); the signal frequency value is obtained using numpy.fft in Python, and the signal average frequency is obtained by statistical analysis of the signal frequency value using the AVERAGE function in Excel; the signal power spectral density is obtained using pwelch() in MATLAB, and the average power spectral density is obtained by statistical analysis of the signal power spectral density using the AVERAGE function in Excel; the signal frequency value, signal average frequency, and average power spectral density are all de-normalized.

[0036] Prior to designing the heart sound data processing method based on time-frequency feature maps provided in this application, a preset database is established to store various set data. The database includes, but is not limited to, autocorrelation values, stationarity test roots, power spectral density, etc., and the various values ​​are directly set by technical personnel. For example, the minimum value of the difference between the autocorrelation values ​​of the collected heart sound signals at a preset historical acquisition monitoring time and the signal at the left adjacent historical acquisition monitoring time represents the reference minimum autocorrelation difference; the result of summing and averaging the stationarity test roots of the collected historical signals represents the reference stationarity test root; and the result of summing and averaging the power spectral density of the collected historical signals represents the reference power spectral density.

[0037] A2, after performing a difference averaging approximation operation on the autocorrelation values ​​of the inner sound signal at a preset time interval and between the preset acquisition and monitoring time and the left adjacent acquisition and monitoring time, the time-domain stationarity test root approximation operation result is subjected to time-domain stationarity interaction processing to obtain the time-domain stationarity. The time-domain stationarity interaction processing describes the interaction between the result of the difference averaging approximation operation on the signal autocorrelation value and the result of the signal stationarity test root approximation operation; specifically, it refers to the process of obtaining the time-domain stationarity after coupling the result of the difference averaging approximation operation on the signal autocorrelation value and the result of the signal stationarity test root approximation operation. The specific constraint expression for the time-domain stationarity is as follows:

[0038]

[0039] In the formula, Let ZR(t) represent the time-domain stationarity, t represent the acquisition and monitoring time of the inner sound signal within a preset time interval, t=1,2,...,T, where T represents the total number of acquisition and monitoring times of the inner sound signal within the preset time interval (T≥2), ZR(t) represent the signal autocorrelation value of the inner sound signal at the t-th acquisition and monitoring time within the preset time interval, ZR(t-1) represent the signal autocorrelation value of the inner sound signal at the (t-1)-th acquisition and monitoring time within the preset time interval, and MIN represent the time-domain stationarity. zc denoted as the reference minimum autocorrelation difference, DG represents the signal stationarity test root, and ΔDG represents the reference stationarity test root.

[0040] A3, after performing difference analysis between the signal frequency values ​​and the signal average frequency, performs frequency domain stationarity interaction processing on the average power spectral density approximation calculation results to obtain the frequency domain stationarity. Frequency domain stationarity interaction processing describes the interaction between the signal frequency value difference analysis results and the average power spectral density approximation calculation results; specifically, it refers to the process of obtaining the frequency domain stationarity after coupling the signal frequency value difference analysis results and the average power spectral density approximation calculation results. The specific constraint expression for frequency domain stationarity is as follows;

[0041]

[0042] In the formula, Let f(t) represent the frequency domain stationarity, and f(t) represent the signal frequency value of the inner sound signal at the t-th acquisition and monitoring time point within a preset time interval. P f P represents the average frequency of the internal sound signal within a preset time interval. GP ΔGP represents the average power spectral density of the internal sound signal within a preset time interval, and ΔGP represents the reference power spectral density.

[0043] A4, time-domain stationarity (i.e. ), frequency domain stationarity (i.e. After weighting and coupling the corresponding stationary compensation quantity, the stationary heart sound segmentation value is obtained. The stationary compensation quantity includes the time-domain stationary compensation quantity (i.e., ) and frequency domain stationary compensation (i.e. The time-domain and frequency-domain stationarity compensation values ​​are used to describe the influence of time-domain and frequency-domain stationarity on the heart sound stationarity segmentation value, respectively. The sum of the time-domain and frequency-domain stationarity compensation values ​​is 1. These values ​​are obtained from a preset database. For example, the real-time time-domain and frequency-domain stationarity values ​​are input into a preset mapping set of time-domain and frequency-domain stationarity values ​​and their corresponding compensation values ​​in the database to obtain the time-domain and frequency-domain stationarity compensation values. The specific constraint expression for the heart sound stationarity segmentation value is as follows:

[0044]

[0045] In the formula, This represents the heart sound steady-state division value. This represents the time-domain stationary compensation amount. This represents the frequency domain stationary compensation amount.

[0046] In summary, the heart sound stability partition value represents the quantitative data on the evaluation of the stability of the heart sound signal characteristics in the time-frequency feature map by the combined effects of time-domain stability and frequency-domain stability. The heart sound stability partition value has an impact on the evaluation of the stability of the heart sound signal characteristics in the time-frequency feature map by the combined effects of time-domain stability and frequency-domain stability. Specifically, as the time-domain stability and frequency-domain stability increase, the heart sound stability partition value also increases.

[0047] In this embodiment, the heart sound stationarity partition value is used to quantitatively evaluate the stationarity of the heart sound signal characteristics in the time-frequency feature map. The heart sound stationarity partition value is obtained through quantitative analysis of the correlation between various parameters. These parameters are interdependent and do not exist independently. For example, the signal autocorrelation value reflects the correlation of the heart sound signal at different time delays and is used to describe the temporal structure of the heart sound signal. Difference analysis of the signal autocorrelation value reveals whether the heart sound signal exhibits periodic or trend-like changes. The signal stationarity test root is used to determine whether the statistical characteristics of the heart sound signal change over time. As the deviation between the calculated autocorrelation difference and the reference minimum autocorrelation difference decreases, it indicates that the temporal characteristics of the heart sound signal are relatively stable. This means the deviation between the signal stationarity test root and the reference stationarity test root also decreases, leading to an increase in time-domain stationarity. Furthermore, as the frequency difference analysis results increase, indicating increased frequency fluctuations and a wider frequency spectrum distribution of the heart sound signal, the power spectrum may exhibit a more pronounced energy distribution in the high-frequency range. This means the deviation between the average power spectral density and the reference power spectral density of the heart sound signal increases, further reducing frequency-domain stationarity. Simultaneously, time-domain and frequency-domain stationarity are interrelated. Higher temporal stationarity in a heart sound signal usually implies relatively stable frequency characteristics. Conversely, if the signal changes drastically over time (e.g., large autocorrelation fluctuations), its frequency fluctuations and power spectral density may also be unstable, resulting in lower frequency-domain stationarity. Therefore, as time-domain stationarity increases, frequency-domain stationarity also increases, leading to an increase in the heart sound stationarity resolution value.

[0048] By quantitatively evaluating the stationarity of the heart sound signal characteristics in the time-frequency feature map, a more accurate evaluation of the stationarity of the heart sound signal characteristics is achieved, thereby enabling a more accurate classification of the differences in the stationarity of the heart sound signal in the identification of heart sound noise.

[0049] Furthermore, the results of the heart sound stability difference characteristics were obtained, and the specific process is as follows:

[0050] B1. Based on the preset heart sound stability characteristic threshold range obtained from the preset database, determine whether the obtained heart sound stability division value is within the preset heart sound stability characteristic threshold range; the preset heart sound stability characteristic threshold range is set by professionals according to the standards in the field.

[0051] B2. If the heart sound stability segmentation value is within the preset heart sound stability characteristic threshold range, then the heart sound stability difference characteristic result is recorded as a stable heart sound signal.

[0052] B3. If the heart sound stability division value is not within the preset heart sound stability characteristic threshold range, then the heart sound stability difference characteristic result is recorded as a non-stationary heart sound signal.

[0053] In this embodiment, the heart sound stability difference characteristic results include stable heart sound signals and non-stable heart sound signals; by combining the preset heart sound stability characteristic threshold range to judge the feature stability of the heart sound signal in the time-frequency feature map, the heart sound stability difference characteristic results after the feature stability difference classification judgment of the heart sound signal are realized, thereby improving the reliability of heart sound signal feature difference processing in the process of heart sound signal noise recognition.

[0054] Furthermore, the specific steps for optimizing heart sounds based on the results of heart sound stability difference characteristics are as follows:

[0055] C1, if the result of the heart sound stability difference characteristic is a stable heart sound signal, then the stable heart sound signal is subjected to stable heart sound optimization processing; stable heart sound optimization processing means optimizing the input of the stable heart sound signal based on the result of the heart sound stability difference characteristic to improve the accuracy of the heart sound signal in the noise recognition process.

[0056] It should be added that the specific process for optimizing steady heart sound signals is as follows:

[0057] C11 divides the steady heart sound signal into short-time steady heart sound signals with equal lengths of steady time window. Based on the overlap between the short-time steady heart sound signal and the reference known noise signal, a matching detection is performed to obtain the overlap matching detection value.

[0058] The method involves matching and detecting overlap between short-time stationary heart sound signals and known reference murmur signals to obtain overlap matching detection values. The specific steps are as follows:

[0059] C111 acquires the overlap detection quantization data between a short-time stationary heart sound signal and a known reference murmur signal. The overlap detection quantization data includes the detection cross-correlation coefficient, the detection similarity coefficient, and the spectral overlap coefficient. Specifically, the detection cross-correlation coefficient is obtained using the xcorr function in MATLAB; the detection similarity coefficient is obtained using the cosine function in MATLAB; and the spectral overlap coefficient is obtained using the pwelch function in MATLAB.

[0060] Specifically, the cross-correlation coefficient is used to describe the similarity between the short-time stationary heart sound signal and the reference known noise signal in the time domain, the similarity coefficient is used to describe the structural similarity between the short-time stationary heart sound signal and the reference known noise signal, and the spectral overlap coefficient is used to describe the similarity between the short-time stationary heart sound signal and the reference known noise signal in the frequency domain.

[0061] C112 involves weighting and coupling the overlap detection quantization data with the corresponding detection compensation value to obtain the overlap matching detection value. The specific constraint expression for the overlap matching detection value is as follows;

[0062]

[0063] In the formula, δ1 represents the cross-correlation detection compensation amount, δ2 represents the similarity detection compensation amount, and δ3 represents the spectral overlap compensation amount. This indicates the cross-correlation coefficient. This represents the similarity coefficient detected. This represents the spectral overlap coefficient.

[0064] The detection compensation amount includes cross-correlation detection compensation amount, similarity detection compensation amount, and spectral overlap compensation amount. The cross-correlation detection compensation amount, similarity detection compensation amount, and spectral overlap compensation amount are obtained by inputting the real-time detection cross-correlation number, detection similarity coefficient, and spectral overlap coefficient into the database and the preset mapping set of detection cross-correlation number, detection similarity coefficient, and spectral overlap coefficient and their corresponding compensation amount.

[0065] The overlap matching detection value represents the quantized data of the overlap detection quantization data for matching the overlap between a short-time stationary heart sound signal and a reference known noise signal. It is used to quantify the degree of overlap between the short-time stationary heart sound signal and the reference known noise signal. The overlap matching detection value includes several parameters. Specifically, as the detection cross-correlation coefficient, detection similarity coefficient, and spectral overlap coefficient increase, the overlap matching detection value increases, indicating an increased degree of overlap between the short-time stationary heart sound signal and the reference known noise signal. Furthermore, the parameters in the overlap matching detection value are not independent. For example, the detection cross-correlation coefficient describes the similarity between the short-time stationary heart sound signal and the reference known noise signal in the time domain. If the signals are well aligned in the time domain, i.e., the detection cross-correlation coefficient increases accordingly, their morphology and structure may also have a high degree of similarity, i.e., the detection similarity coefficient also increases accordingly. Simultaneously, the spectral overlap coefficient describes the similarity between the short-time stationary heart sound signal and the reference known noise signal in the frequency domain. When two signals have strong overlap in the frequency domain, their time-domain waveforms are usually also similar, especially in periodic signals. Therefore, the spectral overlap coefficient is also interrelated with the detection cross-correlation coefficient and the detection similarity coefficient, which together quantify the degree of matching between the detected short-time stationary heart sound signal and the reference known noise signal.

[0066] The steady-state time window length represents the result of mapping the deviation between the real-time steady-state heart sound segmentation value and the steady-state heart sound reference threshold to a steady-state segmentation length mapping set in a preset database. The steady-state segmentation length mapping set represents the mapping relationship between the steady-state heart sound segmentation value, the deviation between the steady-state heart sound reference threshold and the steady-state time window length.

[0067] C12. If the deviation between the overlap match detection value and the reference overlap match threshold is within a preset allowable deviation range obtained from the preset database, then the input confidence scores of the short-time stable heart sound signals are sorted in descending order to obtain the murmur level recognition input priority for the short-time stable heart sound signals. Based on the murmur level recognition input priority, a preset number of short-time stable heart sound signals are input into the constructed murmur level recognition model to obtain the murmur level recognition result. The reference overlap match threshold is the result of summing and averaging the collected historical overlap match detection values. The preset allowable deviation range is set by professionals according to industry standards; for example, the preset allowable deviation range is set to be 3.0 to 5.0.

[0068] The input confidence score involved represents the result of mapping the deviation between the real-time overlapping match detection value and the reference overlapping match threshold to an input confidence mapping set in a preset database. The input confidence mapping set represents the mapping relationship between the deviation between the overlapping match detection value and the reference overlapping match threshold and the input confidence score. As the deviation between the overlapping match detection value and the reference overlapping match threshold increases, the input confidence score decreases accordingly.

[0069] C13. If the deviation between the overlap matching detection value and the reference overlap matching threshold is not within the preset allowable deviation range obtained from the preset database, then principal component analysis is performed based on the feature dimensionality reduction strength of the short-time stationary heart sound signal. The preset number of short-time stationary heart sound signals after principal component analysis dimensionality reduction are input into the constructed murmur level recognition model to obtain the murmur level recognition result.

[0070] The feature dimensionality reduction intensity involved represents the result of mapping the deviation between the real-time overlapping match detection value and the reference overlapping match threshold to a dimensionality reduction intensity mapping set in a preset database. The dimensionality reduction intensity mapping set represents the mapping relationship between the deviation between the overlapping match detection value and the reference overlapping match threshold and the feature dimensionality reduction intensity. As the deviation between the overlapping match detection value and the reference overlapping match threshold increases, the feature dimensionality reduction intensity also increases.

[0071] C2, if the result of the heart sound stationary difference characteristic is a non-stationary heart sound signal, then the non-stationary heart sound signal is subjected to non-stationary heart sound optimization processing; non-stationary heart sound optimization processing means optimizing the input of non-stationary heart sound signal based on the result of the heart sound stationary difference characteristic to improve the accuracy of the heart sound signal in the process of noise recognition.

[0072] It should be added that the specific process for optimizing non-stationary heart sound signals is as follows:

[0073] C21, determine whether the deviation between the heart sound stability segmentation value and the heart sound stability reference threshold is greater than the stability deviation setting value obtained from the preset database; the deviation between the heart sound stability segmentation value and the heart sound stability reference threshold is the result of taking the absolute value of the difference between the heart sound stability segmentation value and the heart sound stability reference threshold; the heart sound stability reference threshold and the stability deviation setting value are obtained from the preset database. For example, the heart sound stability reference threshold is set to the maximum value of the preset heart sound stability characteristic threshold range, and the stability deviation setting value is the maximum value of the deviation between the collected historical heart sound stability segmentation values ​​and the heart sound stability reference threshold.

[0074] C22, if the deviation between the heart sound stability division value and the heart sound stability reference threshold is greater than the stability deviation setting value obtained from the preset database, then the deviation between the heart sound stability division value and the stability deviation setting value is input into the heart sound sampling mapping set in the preset database to obtain the heart sound signal sampling frequency. The heart sound signal is collected at the heart sound signal sampling frequency. The heart sound sampling mapping set represents the mapping relationship between the deviation between the heart sound stability division value and the stability deviation setting value and the heart sound signal sampling frequency. As the deviation between the heart sound stability division value and the stability deviation setting value increases, the heart sound signal sampling frequency increases accordingly.

[0075] C23, if the deviation between the stable heart sound segmentation value and the stable heart sound reference threshold is not greater than the stable deviation setting value obtained from the preset database, then the non-stationary heart sound signal is divided into short-time non-stationary heart sound signals with the length of the non-stationary time window.

[0076] The non-stationary time window length involved represents the result of mapping the deviation between the real-time stationary heart sound segmentation value and the stationary heart sound reference threshold into a non-stationary segmentation length mapping set in a preset database. The non-stationary segmentation length mapping set represents the mapping relationship between the deviation between the stationary heart sound segmentation value and the stationary heart sound reference threshold and the non-stationary time window length. As the deviation between the stationary heart sound segmentation value and the stationary heart sound reference threshold increases, the non-stationary time window length decreases accordingly.

[0077] C24 retrieves the signal energy value of a short-term non-stationary heart sound signal and determines whether this value is greater than a reference signal energy threshold obtained from a preset database. If the signal energy value is greater than the reference threshold, the short-term non-stationary heart sound signal is marked as a valid heart sound signal; otherwise, it is marked as a waiting heart sound signal. The signal energy value is retrieved using a signal processing tool (such as LabVIEW), and the reference signal energy threshold is represented by the summation and averaging of collected historical signal energy values.

[0078] C25 determines whether all short-term non-stationary heart sound signals within the adjacent non-stationary time window of the valid heart sound signal are valid heart sound signals. The length of the adjacent non-stationary time window includes the left and right adjacent non-stationary time windows of the valid heart sound signal. If all short-term non-stationary heart sound signals within the adjacent non-stationary time window length of the valid heart sound signal are valid heart sound signals, then the valid heart sound signal is marked as a valid priority heart sound signal; otherwise, no marking is performed.

[0079] C26 inputs the effective priority heart sound signal into the constructed murmur level recognition model to obtain the murmur level recognition result.

[0080] In this embodiment, the noise level recognition model incorporates a three-layer attention mechanism: spatial attention, channel attention, and frame attention. Spatial attention is first embedded in each scale branch of the multi-scale feature extraction network. Its purpose is to enable the model to adaptively focus on highly discriminative regions in the Mel spectrum, assigning these regions higher weights when extracting noise, reducing interference from irrelevant regions, and improving feature discriminativity. Then, features from each scale are concatenated using channel stacking and aggregated with different weights using channel attention to obtain the noise level for a single cardiac cycle. This aggregation method can, to some extent, play a scale selection role, enabling more accurate extraction of noise features from the Mel spectrum. Finally, the frame attention mechanism, when aggregating features from different cardiac cycles, adaptively determines the aggregation coefficients based on the complementary and redundant characteristics between cardiac cycles, reducing the impact of differences between different cardiac cycles on noise features and improving feature invariance. The three levels of attention mechanisms are closely integrated; the heart sound optimization processing includes steady heart sound optimization processing and non-steady heart sound optimization processing; by performing corresponding heart sound optimization processing when the heart sound steady difference characteristic result is a steady heart sound signal or a non-steady heart sound signal, the difference optimization processing of the heart sound signal input for noise recognition is realized by combining the steady difference of the heart sound signal, thereby improving the input accuracy of the heart sound signal in the noise recognition process.

[0081] Furthermore, the noise level identification results are verified, and the specific process is as follows:

[0082] D1. If the heart sound stability difference characteristic result is a stable heart sound signal, then input a preset multiple of short-time stable heart sound signals into the constructed murmur level recognition model to obtain the verification murmur level recognition result; wherein, the preset multiple is set by professionals according to the standards in the field, for example, the preset multiple is set to be one times the number of short-time stable heart sound signals input.

[0083] D2. If the heart sound stability difference characteristic result is a non-stationary heart sound signal, then input a preset multiple of valid heart sound signals into the constructed noise level recognition model to obtain the verification noise level recognition result.

[0084] D3. Determine if the conformity between the noise level identification result and the verification noise level identification result is greater than a preset conformity threshold obtained from the preset database. If the conformity is greater than the preset threshold, the verification is successful; otherwise, update the noise level identification result. The conformity is the ratio of the total number of times the noise level identification result matches the verification result to the total number of times the noise level identification result is verified. The preset conformity threshold is set by professionals according to industry standards; for example, a preset conformity threshold of 0.8.

[0085] It should be added that the specific steps for updating the noise level identification results are as follows:

[0086] D31 verifies the noise level identification result. If the verification fails, the output result is verified sequentially. If the verification is successful, the noise level identification result is updated to the output result that was verified.

[0087] D32: If verification fails even after the number of input heart sound signals reaches the preset maximum number, the noise level identification result will be updated to "identification error". The preset maximum number is set by professionals according to industry standards; for example, the preset maximum number may be set to five times the number of input heart sound signals.

[0088] In this embodiment, the noise level identification result is verified and updated by combining the conformity of the noise level identification result with the verification noise level identification result with a preset conformity threshold. This realizes the verification and updating of the noise level identification result obtained during the noise identification process of heart sound signals, thereby improving the reliability of the noise level identification result obtained during the noise identification process of heart sound signals.

[0089] like Figure 3 The diagram shown is a structural schematic of a heart sound data processing system based on a time-frequency feature map provided in this application embodiment. The heart sound data processing system based on a time-frequency feature map provided in this application embodiment includes: a heart sound stationarity segmentation module, an optimization recognition processing module, and a noise recognition verification module. The heart sound stationarity segmentation module is used to acquire heart sound signals within a preset time interval, perform feature stationarity difference segmentation on the time-frequency feature map obtained based on the heart sound signals, and obtain heart sound stationarity difference characteristic results. The optimization recognition processing module is used to perform heart sound optimization processing based on the heart sound stationarity difference characteristic results to obtain noise level recognition results. Heart sound optimization processing means optimizing the heart sound signal input during the noise recognition process based on the heart sound stationarity difference characteristic results to improve the accuracy of heart sound signal noise recognition. The noise recognition verification module is used to verify the noise level recognition results. If the verification is successful, the noise level recognition results are fed back; otherwise, the noise level recognition results are updated.

[0090] In summary, this application embodiment obtains the stationary difference characteristics of heart sounds by dividing the time-frequency feature map obtained based on the heart sound signal into stationary differences. Then, it performs heart sound optimization processing based on the stationary difference characteristics and obtains the noise level identification result. Finally, it verifies the noise level identification result. If the verification is successful, the noise level identification result is fed back; otherwise, the noise level identification result is updated. This realizes the analysis of the stationary difference characteristics of the heart sound signal and the optimization of the heart sound signal input in the noise identification process, thereby achieving more accurate noise identification of the heart sound signal. It effectively solves the problem in the prior art that the stationary difference of the heart sound signal is not fully considered in the process of noise identification of the heart sound signal based on the time-frequency feature map.

[0091] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0092] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for processing heart sound data based on time-frequency feature maps, characterized in that, Includes the following steps: S1, acquire the heart sound signal within a preset time interval, perform feature stationary difference division on the time-frequency feature map obtained based on the heart sound signal, and obtain the heart sound stationary difference characteristic result; S2, combine the results of heart sound stability difference characteristics to perform heart sound optimization processing, and obtain the noise level recognition result. The heart sound optimization processing means optimizing the heart sound signal input in the noise recognition process based on the results of heart sound stability difference characteristics to improve the accuracy of heart sound signal noise recognition. S3 verifies the noise level identification result. If the verification is successful, the noise level identification result is fed back; otherwise, the noise level identification result is updated. The specific steps for performing feature stationary difference segmentation on the time-frequency feature map obtained based on heart sound signals are as follows: The feature segmentation data of the inner sound signal at a preset time interval is obtained. The feature segmentation data includes the signal autocorrelation value, the signal stationarity test root, the signal frequency value, the signal average frequency, and the average power spectral density. After performing a difference averaging approximation operation on the autocorrelation values ​​of the inner sound signal at a preset time interval and at the preset acquisition and monitoring time and the left adjacent acquisition and monitoring time, the time domain stationarity test root approximation operation result is subjected to time domain stationarity interaction processing to obtain the time domain stationarity. The time domain stationarity interaction processing is used to describe the interaction between the signal autocorrelation value difference averaging approximation operation result and the signal stationarity test root approximation operation result. After performing a difference analysis between the signal frequency value and the average signal frequency, the average power spectral density approximation calculation result is subjected to frequency domain stationarity interaction processing to obtain the frequency domain stationarity. The frequency domain stationarity interaction processing is used to describe the interaction between the signal frequency value difference analysis result and the average power spectral density approximation calculation result. The time-domain stationarity, frequency-domain stationarity, and corresponding stationarity compensation are weighted and coupled to obtain the heart sound stationarity partition value. The heart sound stationarity partition value is used to quantitatively evaluate the stationarity of the heart sound signal features in the time-frequency feature map. The stationary compensation amount includes time-domain stationary compensation amount and frequency-domain stationary compensation amount; The heart sound stability division value represents the quantitative data that evaluates the stability of the heart sound signal characteristics in the time-frequency feature map by combining time-domain stability and frequency-domain stability. The specific process for obtaining the results of the heart sound stability difference characteristics is as follows: Based on the preset heart sound stability characteristic threshold range obtained from the preset database, determine whether the obtained heart sound stability division value is within the preset heart sound stability characteristic threshold range. If the heart sound stability segmentation value is within the preset heart sound stability characteristic threshold range, then the heart sound stability difference characteristic result is recorded as a stable heart sound signal. If the heart sound stability division value is not within the preset heart sound stability characteristic threshold range, the heart sound stability difference characteristic result is recorded as a non-stationary heart sound signal. The heart sound stability difference characteristics results include stable heart sound signals and non-stable heart sound signals.

2. The method for processing heart sound data based on time-frequency feature maps as described in claim 1, characterized in that, The specific steps for optimizing heart sounds by combining the results of heart sound stability difference characteristics are as follows: If the heart sound stability difference characteristic result is a stable heart sound signal, then the stable heart sound signal is subjected to stable heart sound optimization processing; The smooth heart sound optimization processing refers to optimizing the smooth heart sound signal input based on the results of the heart sound smoothness difference characteristics to improve the accuracy of the heart sound signal in the noise recognition process. If the heart sound stability difference characteristic result is a non-stationary heart sound signal, then the non-stationary heart sound signal is subjected to non-stationary heart sound optimization processing; The non-stationary heart sound optimization processing refers to optimizing the input of non-stationary heart sound signals based on the results of heart sound stationary difference characteristics to improve the accuracy of heart sound signal noise recognition process; The heart sound optimization processing includes steady heart sound optimization processing and non-steady heart sound optimization processing.

3. The method for processing heart sound data based on time-frequency feature maps as described in claim 2, characterized in that, The specific process for optimizing the steady heart sound signal is as follows: The steady heart sound signal is divided into short-time steady heart sound signals with equal length of steady time window. The overlap between the short-time steady heart sound signal and the reference known noise signal is matched and detected to obtain the overlap matching detection value. The overlap matching detection value is used to quantify the degree of matching between the short-time steady heart sound signal and the reference known noise signal. The steady-state time window length represents the result of mapping the deviation between the real-time steady-state heart sound segmentation value and the steady-state heart sound reference threshold to a steady-state segmentation length mapping set in a preset database. The steady-state segmentation length mapping set represents the mapping relationship between the steady-state heart sound segmentation value, the deviation between the steady-state heart sound reference threshold and the steady-state time window length. If the deviation between the overlap matching detection value and the reference overlap matching threshold is within the preset allowable deviation range obtained from the preset database, the noise level recognition input priority of the short-time stationary heart sound signal is obtained by sorting the input confidence score of the short-time stationary heart sound signal in descending order. According to the noise level recognition input priority, a preset number of short-time stationary heart sound signals are input into the constructed noise level recognition model to obtain the noise level recognition result. The input confidence score represents the result of mapping the deviation between the real-time overlapping match detection value and the reference overlapping match threshold to an input confidence mapping set in a preset database. The input confidence mapping set represents the mapping relationship between the deviation between the overlapping match detection value and the reference overlapping match threshold and the input confidence score. If the deviation between the overlap matching detection value and the reference overlap matching threshold is not within the preset allowable deviation range obtained from the preset database, then principal component analysis is performed based on the feature dimensionality reduction strength of the short-time stationary heart sound signal. The preset number of short-time stationary heart sound signals after principal component analysis dimensionality reduction are input into the constructed murmur level recognition model to obtain the murmur level recognition result. The feature dimensionality reduction strength represents the result of mapping the deviation between the real-time overlapping match detection value and the reference overlapping match threshold to a dimensionality reduction strength mapping set in a preset database. The dimensionality reduction strength mapping set represents the mapping relationship between the deviation between the overlapping match detection value and the reference overlapping match threshold and the feature dimensionality reduction strength.

4. The method for processing heart sound data based on time-frequency feature maps as described in claim 3, characterized in that, The method of matching and detecting overlap between short-time stationary heart sound signals and known reference murmur signals to obtain overlap matching detection values ​​involves the following steps: Acquire overlap detection quantization data between a short-time stationary heart sound signal and a reference known murmur signal, wherein the overlap detection quantization data includes a detection cross-correlation coefficient, a detection similarity coefficient, and a spectral overlap coefficient; The detection cross-correlation coefficient is used to describe the similarity between the short-time stationary heart sound signal and the reference known noise signal in the time domain; the detection similarity coefficient is used to describe the structural similarity between the short-time stationary heart sound signal and the reference known noise signal; and the spectral overlap coefficient is used to describe the similarity between the short-time stationary heart sound signal and the reference known noise signal in the frequency domain. The overlapping detection quantization data and the corresponding detection compensation amount are weighted and coupled to obtain the overlapping matching detection value. The detection compensation amount includes cross-correlation detection compensation amount, similarity detection compensation amount and spectral overlap compensation amount. The overlap matching detection value represents the quantized data of the overlap detection quantization data for the overlap detection of short-time stationary heart sound signals and reference known noise signals.

5. The method for processing heart sound data based on time-frequency feature maps as described in claim 2, characterized in that, The specific process for optimizing non-stationary heart sound signals is as follows: Determine whether the deviation between the heart sound stability classification value and the heart sound stability reference threshold is greater than the stability deviation setting value obtained from the preset database; If the deviation between the heart sound stability division value and the heart sound stability reference threshold is greater than the stability deviation setting value obtained from the preset database, then the deviation between the heart sound stability division value and the stability deviation setting value is input into the heart sound sampling mapping set in the preset database to obtain the heart sound signal sampling frequency. The heart sound signal is then collected using the heart sound signal sampling frequency. The heart sound sampling mapping set represents the mapping relationship between the deviation between the heart sound stability division value and the stability deviation setting value and the heart sound signal sampling frequency. If the deviation between the stable heart sound segmentation value and the stable heart sound reference threshold is not greater than the stable deviation setting value obtained from the preset database, then the non-stationary heart sound signal is divided into short-time non-stationary heart sound signals with the length of the non-stationary time window. The non-stationary time window length represents the result of mapping the deviation between the real-time stable heart sound segmentation value and the stable heart sound reference threshold to a non-stationary segmentation length mapping set in a preset database. The non-stationary segmentation length mapping set represents the mapping relationship between the stable heart sound segmentation value, the deviation between the stable heart sound reference threshold and the non-stationary time window length. Obtain the signal energy value of short-term non-stationary heart sound signals and determine whether the signal energy value of the short-term non-stationary heart sound signals is greater than the reference signal energy threshold obtained from the preset database; If the signal energy value of a short-term non-stationary heart sound signal is greater than the reference signal energy threshold, the short-term non-stationary heart sound signal is marked as a valid heart sound signal; otherwise, the short-term non-stationary heart sound signal is marked as a waiting heart sound signal. Determine whether all short-term non-stationary heart sound signals in adjacent non-stationary time windows of a valid heart sound signal are valid heart sound signals. The length of the adjacent non-stationary time window includes the left adjacent non-stationary time window and the right adjacent non-stationary time window of the valid heart sound signal. If all short-time non-stationary heart sound signals within the adjacent non-stationary time window length of a valid heart sound signal are valid heart sound signals, then the valid heart sound signal is marked as a valid priority heart sound signal; otherwise, it is not marked. The effective priority heart sound signal is input into the constructed noise level recognition model to obtain the noise level recognition result.

6. The method for processing heart sound data based on time-frequency feature maps as described in claim 5, characterized in that, The specific process for verifying the noise level identification results is as follows: If the heart sound stability difference characteristic result is a stable heart sound signal, then input a preset multiple of short-time stable heart sound signals into the constructed murmur level recognition model to obtain the verification murmur level recognition result; If the heart sound stability difference characteristic result is a non-stationary heart sound signal, then input a preset multiple of valid heart sound signals into the constructed murmur level recognition model to obtain the verification murmur level recognition result; Determine whether the degree of agreement between the noise level identification result and the verification noise level identification result is greater than the preset result agreement threshold obtained from the preset database; If the degree of agreement between the noise level identification result and the verification noise level identification result is greater than the threshold value of the preset result obtained from the preset database, then the verification is successful; otherwise, the noise level identification result is updated.

7. The method for processing heart sound data based on time-frequency feature maps as described in claim 6, characterized in that, The specific steps for updating the noise level identification results are as follows: Verify the noise level identification result. If the verification fails, verify the output result in turn. If the verification is successful, update the noise level identification result to the output result that was verified. If the verification fails even after the number of input heart sound signals reaches the preset maximum number, the noise level recognition result will be updated to recognition anomaly.

8. A heart sound data processing system based on time-frequency feature maps, applied to the heart sound data processing method based on time-frequency feature maps as described in any one of claims 1 to 7, characterized in that, include: The system includes a heart sound smoothing segmentation module, an optimized recognition and processing module, and a noise recognition and verification module. The heart sound stationarity segmentation module is used to acquire heart sound signals within a preset time interval, perform feature stationarity difference segmentation on the time-frequency feature map obtained based on the heart sound signals, and obtain heart sound stationarity difference characteristic results. The optimized recognition processing module is used to perform heart sound optimization processing based on the heart sound stability difference characteristics to obtain the murmur level recognition result. The heart sound optimization processing means optimizing the heart sound signal input in the murmur recognition process based on the heart sound stability difference characteristics to improve the accuracy of heart sound signal murmur recognition. The noise recognition and verification module is used to verify the noise level recognition result. If the verification is successful, the noise level recognition result is fed back; otherwise, the noise level recognition result is updated.

Images