A method and system for self-monitoring of infant asphyxia risk based on laryngeal breath sounds

Abstract

The present application relates to the technical field of health monitoring, in particular to a kind of infant asphyxia risk autonomous monitoring method and system based on laryngeal respiratory sound, comprising: utilize high sensitivity MEMS microphone to collect the airflow and respiratory sound signal of infant larynx, and utilize micro accelerometer to monitor the thoracic movement data of infant synchronously;The airflow and respiratory sound signal are denoising, and the denoising respiratory sound signal is obtained;The denoising respiratory sound signal is separated according to respiratory event Signal, and is enhanced and combined into respiratory sound data set;Respiratory sound feature is extracted in respiratory sound data set, and asphyxia anomaly classification model is constructed based on respiratory sound feature by machine learning model, and dynamic baseline model that adapts individual respiratory difference is established according to respiratory sound feature.The present application utilizes the hardware device of high sensitivity MEMS microphone and micro accelerometer and artificial intelligence model to combine, realizes the use of laryngeal respiratory sound signal in home environment and autonomously completes infant asphyxia risk identification and early warning.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of health monitoring, in particular to a self-monitoring method and system for infant asphyxia risk based on laryngeal respiratory sound. BACKGROUND

[0002] Infant asphyxia is the main inducement of sudden infant death syndrome (SIDS) and acute accidental events, especially during sleep or feeding, the risk of asphyxia caused by choking, respiratory obstruction or improper body position is extremely high. For these high-risk events of asphyxia, there is currently a lack of continuous, non-invasive and low-cost home monitoring methods. Although hospital-level polysomnography is accurate, it is not suitable for routine use in home settings due to its complex equipment and numerous electrodes; wearable heart rate and oxygen saturation solutions are susceptible to motion and skin temperature interference, resulting in high false positive rates. In recent years, intelligent cameras based on crying or body movement have emerged, but they are not sensitive to silent airway obstruction and have significant privacy concerns. Laryngeal respiratory sound directly reflects the patency of the upper airway, and its spectral characteristics appear wheezing, weakening or disappearing in the early stage of asphyxia, but existing stethoscopes and acoustic patches lack real-time analysis algorithms for infants, making it difficult to achieve closed-loop interaction with care terminals. Therefore, there is an urgent need for a miniature system and method that can autonomously identify and warn of infant asphyxia risk using laryngeal respiratory sound signals in a home environment. SUMMARY

[0003] The present application aims to provide a self-monitoring method for infant asphyxia risk based on laryngeal respiratory sound, to solve the technical problem that existing technologies lack real-time analysis algorithms for infants, making it difficult to autonomously identify and warn of infant asphyxia risk in a home environment.

[0004] To solve the above technical problems, the present application specifically provides the following technical solutions:

[0005] A self-monitoring method for infant asphyxia risk based on laryngeal respiratory sound, comprising the following steps:

[0006] Collecting airflow and respiratory sound signals from the infant's larynx using a high-sensitivity MEMS microphone, and synchronously monitoring thoracic movement data of the infant using a miniature accelerometer;

[0007] Integrating the airflow and respiratory sound signals with the thoracic movement data, and performing noise reduction processing on the airflow and respiratory sound signals to obtain a denoised respiratory sound signal;

[0008] Signal separation of the denoised respiratory sound signal according to respiratory events, and enhancement and combination into a respiratory sound data set;

[0009] Respiratory sound features are extracted from the respiratory sound data set, and a normal, low-risk and high-risk asphyxia anomaly classification model is constructed based on the respiratory sound features through a machine learning model, and a dynamic baseline model that adapts to individual respiratory differences is established according to the respiratory sound features;

[0010] The dynamic baseline model is used for monitoring and warning the respiratory abnormalities of the infant, and the asphyxia anomaly classification model is used for classifying the respiratory abnormalities fed back by the dynamic baseline model.

[0011] As a preferred scheme of the present application, the method for integrating the airflow and respiratory sound signals and the thoracic movement data, and performing noise reduction processing on the airflow and respiratory sound signals comprises:

[0012] The airflow and respiratory sound signals are decomposed, and the airflow and respiratory sound signals are: , wherein, is the airflow and respiratory sound signal, is the target respiratory sound signal, is a noise signal generated by body movement, is an environmental noise signal in

[0013] The thoracic movement data is decomposed, and the thoracic movement data is: , wherein, is the thoracic movement data, is a noise signal generated by body movement, is an environmental noise signal in

[0014] The iterative operation of the least mean square algorithm LMS makes the adaptive filter output noise estimation , and the target respiratory sound signal is calculated according to the noise estimation, wherein, is the noise estimation value, is the set of thoracic movement data, , is the thoracic movement data at the previous moment, is the thoracic movement data at the previous L-1 moments, is the filter weight, is the target respiratory sound signal;

[0015] The target respiratory sound signal is taken as the denoised respiratory sound signal.

[0016] As a preferred scheme of the present application, the method for carrying out signal separation on the denoised respiratory sound signal according to respiratory events comprises:

[0017] The denoised respiratory sound signal is framed into a plurality of audio samples according to a preset time length, and the audio samples less than the preset time length are length-completed;

[0018] The short-time energy (STE) and the short-time zero-crossing rate (ZCR) of each audio sample are calculated, and the audio samples without sound are removed based on the short-time energy (STE) and the short-time zero-crossing rate (ZCR) through a double-threshold method;

[0019] In the remaining audio samples, the audio samples corresponding to the respiratory events of inhalation, exhalation, crying, coughing and abnormal respiratory sound are identified based on the short-time energy (STE) and the short-time zero-crossing rate (ZCR) and reserved as value-type samples.

[0020] As a preferred scheme of the present application, the method for combining the value-type samples into a respiratory sound dataset comprises:

[0021] Time domain displacement: shifting the value-type samples backward by 1s on the time axis, and splicing the 7.0s to 8.0s part of the value-type samples to the front end of the value-type samples;

[0022] Gain adjustment: adjusting the amplitude of each value-type sample to be lower or higher, wherein the oscillation parameter of the amplitude changed by the adjustment is a random value between 0.1 and 5.0;

[0023] Noise introduction: adding a random noise signal to the value-type samples, and adjusting the signal-to-noise ratio to a preset value;

[0024] Mixing the value-type samples and the enhanced value-type samples as the respiratory sound dataset.

[0025] As a preferred scheme of the present application, the method for extracting respiratory sound features comprises:

[0026] The samples in the respiratory sound dataset are sequentially subjected to pre-emphasis, framing, windowing, fast Fourier transform, mel filter bank, logarithmic energy processing to obtain a mel spectrum graph, and the mel spectrum graph is subjected to DCT transform processing to obtain MFCC features;

[0027] The samples in the respiratory sound dataset are estimated for fundamental frequency features through a cepstrum method, and harmonic energy features are calculated based on the fundamental frequency features;

[0028] The samples in the respiratory sound dataset are sequentially calculated for exhalation time length and amplitude, and inhalation time length and amplitude;

[0029] The MFCC features, the fundamental frequency features, the harmonic energy features, the exhalation time length and amplitude, and the inhalation time length and amplitude are taken as respiratory sound features.

[0030] As a preferred scheme of the present application, the method for constructing the suffocation anomaly classification model comprises:

[0031] The samples in the respiratory sound data set are type-labeled according to normal, low risk and high risk, and the labeled respiratory sound data set is randomly divided into a training set and a test set;

[0032] In the training set, the respiratory sound features are taken as input, and the normal, low risk and high risk are taken as output to train the CNN algorithm architecture, so as to obtain the suffocation anomaly classification model;

[0033] The CNN6 algorithm architecture comprises four convolution blocks ConvBlock, one feature aggregation layer and one fully connected classification layer, and each convolution block adopts a series combination of a convolution layer Conv2D and a maximum pooling layer MaxPooling2D;

[0034] The four convolution blocks are used to gradually extract deep features in the respiratory sound features;

[0035] The feature aggregation layer is used to integrate the deep features to form a global semantic representation;

[0036] The fully connected classification layer is used to output the classification results of normal, low risk and high risk.

[0037] As a preferred scheme of the present application, the method for constructing the suffocation anomaly classification model comprises:

[0038] A normal respiratory sound signal conforming to a specified time length is selected from the denoised respiratory sound signal as a learning period signal;

[0039] The MFCC features, the fundamental frequency features, the harmonic energy features, the expiration time length and amplitude, the inspiration time length and amplitude of the learning period signal are obtained, and the mean and standard deviation of each feature within the specified time length are calculated;

[0040] According to the Gaussian distribution of the features within the specified time length, the baseline model of each feature is established , and is combined as the dynamic baseline model , wherein, is the baseline model of the feature z, is the mean of the feature z, is the standard deviation of the feature z.

[0041] As a preferred scheme of the present application, the method for monitoring and warning the respiratory anomaly of an infant by using the dynamic baseline model and classifying the suffocation risk of the respiratory anomaly fed back by the dynamic baseline model by using the suffocation anomaly classification model comprises:

[0042] The dynamic baseline model is used to monitor the deviation degree of the MFCC features, the fundamental frequency features, the harmonic energy features, the expiration time length and amplitude, and the inspiration time length and amplitude of the real-time denoised breathing sound signal of the infant, wherein:

[0043] When the MFCC features, the fundamental frequency features, the harmonic energy features, the expiration time length and amplitude, and the inspiration time length and amplitude of the real-time denoised breathing sound signal of the infant deviate from the dynamic baseline model by more than a preset threshold, it is determined that the infant is in a respiratory abnormal state;

[0044] When the MFCC features, the fundamental frequency features, the harmonic energy features, the expiration time length and amplitude, and the inspiration time length and amplitude of the real-time denoised breathing sound signal of the infant deviate from the dynamic baseline model by less than a preset threshold, it is determined that the infant is not in a respiratory abnormal state;

[0045] The MFCC features, the fundamental frequency features, the harmonic energy features, the expiration time length and amplitude, and the inspiration time length and amplitude of the denoised breathing sound signal corresponding to the infant in the respiratory abnormal state are input into the asphyxia abnormality classification model to determine the asphyxia risk category of the infant.

[0046] As a preferred scheme of the present application, the present application provides an infant asphyxia risk autonomous monitoring system based on throat breathing sound, which is applied to an infant asphyxia risk autonomous monitoring method based on throat breathing sound, comprising:

[0047] A data acquisition unit comprising a high-sensitivity MEMS microphone and a micro accelerometer, wherein the high-sensitivity MEMS microphone is used to acquire airflow and breathing sound signals of the throat of the infant, and the micro accelerometer is used to synchronously monitor the thoracic movement data of the infant;

[0048] A signal denoising unit is used to denoise the airflow and breathing sound signals by integrating the airflow and breathing sound signals and the thoracic movement data to obtain denoised breathing sound signals;

[0049] A machine learning unit is used to separate abnormal breathing sound signals from normal breathing sound signals in the denoised breathing sound signals, and combine the normal breathing sound signals and the abnormal breathing sound signals into a breathing sound data set; breathing sound features are extracted from the breathing sound data set, an asphyxia abnormality classification model distinguishing normal, low-risk and high-risk is constructed based on the breathing sound features through a machine learning model, and a dynamic baseline model adapting to individual respiratory differences is established according to the breathing sound features of the normal breathing sound signals;

[0050] An autonomous monitoring unit is used to monitor and warn the respiratory abnormalities of the infant by using the dynamic baseline model, and to classify the asphyxia risk of the respiratory abnormalities fed back by the dynamic baseline model by using the asphyxia abnormality classification model.

[0051] As a preferred scheme of the present application, the present application provides a computer readable storage medium, wherein computer execution instructions are stored, and when a processor executes the computer execution instructions, a method for autonomous monitoring of infant suffocation risk based on laryngeal breath sounds is realized.

[0052] Compared with the prior art, the present application has the following beneficial effects:

[0053] The present application directly collects laryngeal breath sounds by using a high-sensitivity MEMS microphone, and reduces environmental noise interference and improves early suffocation recognition accuracy through noise reduction processing by a miniature accelerometer.

[0054] The embedded AI algorithm formed by the machine learning model and the dynamic baseline model can effectively distinguish between normal breath variation precursors and suffocation categories, and realize real-time analysis for infants.

[0055] The present application combines the hardware devices of high-sensitivity MEMS microphones and miniature accelerometers with artificial intelligence models to realize autonomous completion of infant suffocation risk identification and early warning in a home environment by using laryngeal breath sound signals. BRIEF DESCRIPTION OF DRAWINGS

[0056] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced below. Obviously, the drawings in the following description are only exemplary, and for those skilled in the art, other drawings can be obtained from the provided drawings without creative labor.

[0057] Figure 1 The flow chart of the method for autonomous monitoring of infant suffocation risk based on laryngeal breath sounds provided for the embodiments of the present application;

[0058] Figure 2 The block diagram of the system for autonomous monitoring of infant suffocation risk based on laryngeal breath sounds provided for the embodiments of the present application;

[0059] Figure 3 The overall architecture diagram of the suffocation risk classification model provided for the embodiments of the present application;

[0060] Figure 4 The loss value change graph and the accuracy rate change curve graph provided for the embodiments of the present application;

[0061] Figure 5 The performance index graph of the suffocation risk classification model provided for the embodiments of the present application. DETAILED DESCRIPTION

[0062] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0063] As shown in the drawings, Figure 1 The present application provides a self-monitoring method for infant suffocation risk based on laryngeal breath sound, comprising the following steps:

[0064] The airflow and breath sound signals of the infant's larynx are collected by using a high-sensitivity MEMS microphone, and the thoracic movement data of the infant is synchronously monitored by using a miniature accelerometer;

[0065] The airflow and breath sound signals are denoised by integrating the airflow and breath sound signals and the thoracic movement data, to obtain denoised breath sound signals;

[0066] The denoised breath sound signals are separated according to the breath events, and are enhanced and combined into a breath sound data set;

[0067] Respiratory sound features are extracted from the breath sound data set, and a normal, low-risk and high-risk suffocation anomaly classification model is constructed based on the respiratory sound features by using a machine learning model, and a dynamic baseline model that adapts to individual respiratory differences is established according to the respiratory sound features;

[0068] The respiratory anomaly of the infant is monitored and warned by using the dynamic baseline model, and the respiratory anomaly fed back by the dynamic baseline model is classified as suffocation risk by using the suffocation anomaly classification model.

[0069] The present application obtains the airflow and breath sound signals of the infant's larynx by using a high-sensitivity MEMS microphone. Since the microphone receives not only breath sound data but also noise generated by body movement and external environmental noise, the present application further adds a miniature accelerometer. The miniature accelerometer only receives noise generated by body movement (which is derived from the same physical movement as the motion noise received by the microphone, and has a high degree of correlation) and external environmental noise, and hardly perceives breath sound. The data received by the miniature accelerometer and the data received by the high-sensitivity MEMS microphone can be integrated by algorithm to retain only breath sound data, thereby achieving the purpose of noise reduction, as follows:

[0070] The method for denoising the airflow and breath sound signals by integrating the airflow and breath sound signals and the thoracic movement data comprises:

[0071] The airflow and breath sound signals are decomposed, and the airflow and breath sound signals are: , wherein, a noise signal resulting from body motion in the airflow and breath sound signal, a target breath sound signal, a noise signal resulting from body motion in the airflow and breath sound signal, a noise signal resulting from body motion in the airflow and breath sound signal, a noise signal resulting from body motion in the airflow and breath sound signal, an ambient noise signal in the airflow and breath sound signal;

[0072] performing a constitutive decomposition of the thorax motion data, wherein, the thorax motion data, a noise signal resulting from body motion in the airflow and breath sound signal, a noise signal resulting from body motion in the airflow and breath sound signal, a noise signal resulting from body motion in the airflow and breath sound signal, an ambient noise signal in the airflow and breath sound signal;

[0073] using an iterative operation of a least mean square algorithm LMS to make the adaptive filter output a noise estimate and to calculate a target breath sound signal from the noise estimate, wherein, the noise estimate, the set of thorax motion data, the thorax motion data of a previous time instant of the set of thorax motion data, the thorax motion data of the previous L-1 time instants of the set of thorax motion data, the filter weight, the target breath sound signal, T being the transpose operator; the target breath sound signal as a denoised breath sound signal.

[0074] The mathematical objective of the LSM algorithm is to minimize the mean square value E[e T (k)] of the output error signal e(k). By continuously adjusting the filter weight w(k) such that y(k) becomes the best estimate of the noise component n(k) in d(k), the denoising purpose is achieved when y(k) ≈ n(k) and e(k) ≈ s(k).

[0075] First, initialize the adaptive filter:

[0076] Set the filter length (order) L (e.g. L = 32 or 64), which determines the complexity of the noise channel that the algorithm can model.

[0077] Initialize the filter weight vector w, usually set to an all-zero vector: w(0) = [0, 0,..., 0] T .

[0078] Initialize the filter weight vector w, usually set to an all-zero vector: w(0) = [0, 0,..., 0] .

[0079] Select a step size parameter μ:

[0080] Where μ is too large: the algorithm converges quickly, but is unstable, and the error will oscillate and even diverge.

[0081] μ is too small: the algorithm is stable, but converges slowly and may not be able to track changing noise.

[0082] Therefore, the present application sets the range of μ as: .

[0083] Then, for each time k, perform the following iterative calculation:

[0084] Step 1: Construct a delay input vector: x_vector(k) = [x(k), x(k-1),..., x(k-L+1)] T This vector is a collection of the current and past L-1 accelerometer signals.

[0085] Step 2: Calculate the output of the adaptive filter, y(k) = w(k) T *x_vector(k) = This y(k) is the noise estimate predicted by the algorithm based on the current weight w(k) and the accelerometer signal x(k).

[0086] Step 3: Calculate the error signal (i.e. the output after noise reduction) e(k) = d(k) - y(k), which is the final purified target respiratory sound signal. And if y(k) perfectly estimates the noise n(k), then e(k) = s(k).

[0087] Step 4: Update the filter weights (LMS core), w(k+1) = w(k) + μ * e(k) * x_vector(k).

[0088] The above is the learning process of the LMS algorithm, which actually means:

[0089] e(k): error measure, if e(k) is large, it means that the current estimate y(k) is poor, and the weight needs to be adjusted vigorously; x_vector(k): update direction, the weight will be adjusted along the direction of the current input signal; μ: learning step size, controls the intensity of adjustment.

[0090] The method for separating the denoised respiratory sound signal according to the respiratory event comprises.

[0091] Frame the denoised respiratory sound signal into multiple audio samples according to a preset time length, and fill in the audio samples that are less than the preset time length to the preset time length;

[0092] Calculate the short-time energy (STE) and the short-time zero-crossing rate (ZCR) of each audio sample, and remove the audio samples without sound based on the short-time energy (STE) and the short-time zero-crossing rate (ZCR) by a double-threshold method.

[0093] In the remaining audio samples, the audio samples corresponding to the breathing event of inhale, exhale, cry, cough and abnormal breath sound are identified based on the short-time energy (STE) and the short-time zero-crossing rate (ZCR) and reserved as valuable samples.

[0094] After obtaining the breathing sound signal (i.e. the denoised breathing sound signal) removed of noise, the present application further separates the same to obtain meaningful or valuable signals in the breathing sound signal, such as the valuable audio signals of exhale, inhale, baby voice, cough, cry, abnormal breath sound and the like, and removes the audio signals without sound and without meaning.

[0095] The short-time energy (STE) measures the amplitude of a signal in a frame. The energy of events such as exhale and cry is high, the energy of inhale is lower, and the energy of a silent section is low.

[0096] The short-time zero-crossing rate (ZCR) measures the number of times a signal crosses zero in a frame. The zero-crossing rate of high-frequency sound (such as some cries and friction sounds) is high, and the zero-crossing rate of low-frequency sound (such as exhale) is low.

[0097] A low energy threshold and a high energy threshold are set.

[0098] If the short-time energy (STE) is greater than the high energy threshold, it is determined that there is sound;

[0099] If the short-time energy (STE) is less than the low energy threshold, it is determined that there is no sound;

[0100] If the short-time energy (STE) is between the high energy threshold and the low energy threshold, the ZCR is combined to determine whether the ZCR is lower than a preset ZCR threshold (indicating low-frequency sound), and it is determined that there is a sound frame, otherwise it is determined that there is no sound frame.

[0101] The determination method of the valuable audio signal is as follows:

[0102] Cough signal: The duration is extremely short, the instantaneous energy is extremely high (the STE is large), the waveform is explosive (rapidly rising and slowly decaying), and the frequency spectrum is wide (the MFCC feature is different from that of breathing).

[0103] Cry / voice signal: The duration is long, there is a stable fundamental frequency and harmonic structure, and the zero-crossing rate is moderate.

[0104] Exhalation signal: moderate duration (e.g. 1-2s), high and stable energy, spectrum concentrated in low frequency, low zero-crossing rate, no stable fundamental frequency.

[0105] Inhalation: short duration (e.g. 0.5-1s), lower energy than exhalation, often accompanied by high-frequency friction sound (leading to an increase in zero-crossing rate).

[0106] The basis of model training is data. Since the data is collected from the real environment, the research object is a baby, and the audio of suffocation condition needs to be collected, the final sample size will inevitably be small. Insufficient sample size will cause the suffocation risk classification model to overfit, making it difficult to learn effective features. In order to make the most of existing data resources, the data length should be reasonably cut first. In order to make each cut sample contain a complete breathing cycle as much as possible, the length of each cut segment is set to 8.0s. The audio samples less than 8s are aligned to 8s using zero padding.

[0107] The present application further obtains a more abundant data set through data enhancement technology. The model trained based on the data set expanded by this technology can reduce the risk of overfitting in feature learning, and ultimately will show more stable classification accuracy and stronger environmental adaptability. Data enhancement is to construct semantically equivalent derivative samples by adding controllable perturbations without destroying the characteristics of the original data, so as to expand the size of the data set. This method enhances the generalization ability of the model by improving the diversity of the data, while keeping the basic features of the data unchanged, as follows:

[0108] The method for combining the value type sample enhancement into the respiratory sound data set comprises:

[0109] Time domain displacement: shift the value type sample backward by 1s on the time axis, and splice the 7.0s to 8.0s part of the value type sample to the front end of the value type sample;

[0110] Gain adjustment: lower or raise the amplitude of each value type sample, wherein the amplitude change of the oscillation parameter is a random value between 0.1 and 5.0;

[0111] Introducing noise: add a random noise signal to the value type sample, and adjust the signal-to-noise ratio to a preset value;

[0112] Mixing the value type sample and the enhanced value type sample as the respiratory sound data set.

[0113] The respiratory sound data set is randomly divided in a ratio of 8:2, wherein 80% of the data is used as a training set for the training of the airway sputum accumulation degree classification model, and the remaining 20% is used as an independent test set to evaluate the performance of the model.

[0114] Although data augmentation effectively improves the diversity of training samples, it also brings significant computational burden. In order to balance the model performance and computational efficiency, feature extraction needs to be performed on the augmented audio data.

[0115] The extraction method of the breath sound features comprises:

[0116] The samples in the breath sound dataset are sequentially pre-emphasized, framed, windowed, fast Fourier transformed, Mel filter banked, and log energy processed to obtain a Mel spectrogram, and the Mel spectrogram is processed by DCT transformation to obtain MFCC features;

[0117] The samples in the breath sound dataset are estimated by the cepstrum method to obtain fundamental frequency features, and the harmonic energy features are calculated based on the fundamental frequency features;

[0118] The samples in the breath sound dataset are sequentially calculated for exhalation duration and amplitude, inhalation duration and amplitude;

[0119] The MFCC features, fundamental frequency features, harmonic energy features, exhalation duration and amplitude, and inhalation duration and amplitude are used as breath sound features.

[0120] The feature extraction technology plays a key role in medical classification and recognition tasks by analyzing the original signal to extract its physically relevant key features to realize signal representation. Through feature dimension reduction, the method can effectively remove signal redundancy and convert high-dimensional data into low-dimensional representation, which not only simplifies the calculation but also improves the recognition performance. The present application adopts Mel-scale Frequency Cepstral Coefficients (MFCC) as one of the extracted features. MFCC is a widely used feature extraction method in speech and audio processing, which does not need to analyze the characteristics of the original signal itself. Compared with traditional features such as short-time energy, time-domain envelope, autocorrelation coefficient, zero-crossing rate, and frequency-domain features such as spectral envelope, spectral flux, and spectral bandwidth, MFCC can better distinguish the subtle differences between different audios, can resist the influence of background noise to a certain extent, and has better robustness. Before extracting MFCC, we first generate a Mel spectrogram, which more intuitively shows the time-frequency structure of the audio signal and is more helpful in capturing acoustic features related to human hearing. The process of extracting MFCC features and Mel spectrogram in the present application is as follows:

[0121] The audio samples in the expanded training set are sequentially pre-emphasized, framed, windowed, fast Fourier transformed, Mel filter banked, and log energy processed to obtain a Mel spectrogram;

[0122] The Mel spectrogram is processed by DCT transformation to obtain MFCC features.

[0123] Pre-emphasis: The first step in the MFCC feature extraction process, its core role is to optimize the spectral characteristics of the speech signal through high-frequency enhancement. Framing: The spectral characteristics of the audio signal change over time, in order to accurately capture the frequency characteristics of different time periods, the pre-emphasized signal needs to be framed to ensure that the spectral distribution within each frame remains relatively stable. Windowing: Using the Hamming window to suppress the sidelobes to improve the continuity of adjacent frames. Short-time Fourier transform: Convert time-domain signals into time-frequency joint representation, and reveal the law of local frequency domain characteristics of signals evolving over time through the energy distribution of time-varying spectrum. Mel filter bank: Its essence is a nonlinear frequency band analysis system that simulates the hearing characteristics of the human ear, composed of a set of overlapping triangular bandpass filters. The filter bank is non-uniformly distributed under the Mel scale, with a dense low-frequency region and a sparse high-frequency region, which conforms to the nonlinear perception characteristics of human ear to pitch. Its role includes: ① Smooth the spectrum and suppress harmonics; ② Enhance the formant characteristics; ③ Improve the low-frequency resolution. This processing makes the spectral features more consistent with human auditory perception, laying the perceptual foundation for speech feature extraction such as MFCC. The log energy is mapped to the Mel scale to form a Mel spectrogram, and then a DCT transform is performed to realize decorrelation, reduce the redundancy between features, avoid performance degradation of the classification model due to data correlation, and make the features more compact and effective.

[0124] In addition to extracting MFCC features, the application also extracts the fundamental frequency and harmonic features in the respiratory sound signal through time-frequency analysis, such as using autocorrelation method or cepstrum method to estimate the fundamental frequency F0, and locating the integer multiple positions (2F0, 3F0,...) of the fundamental frequency F0 on the power spectrum, and calculating the energy sum within the bandwidth near these harmonic frequencies to obtain the harmonic energy.

[0125] In addition to the above high-dimensional feature extraction, the application also extracts low-dimensional features such as exhalation duration and amplitude, inhalation duration and amplitude as respiratory sound features.

[0126] The application obtains a multi-modal respiratory sound feature (combination of high and low dimensional features), and constructs a suffocation abnormality classification model on the multi-modal respiratory sound feature, thereby providing more rich feature information for the model, improving the perception dimension of the model, and thereby ensuring the classification performance of the classification model. The use of fusion features for classification judgment significantly enhances the robustness and anti-interference ability of the model.

[0127] The construction method of the suffocation abnormality classification model includes:

[0128] The samples in the respiratory sound data set are labeled according to normal, low risk and high risk, mainly by judging the respiratory rhythm, respiratory continuity, laryngeal stridor degree, cough, and whether there are abnormal respiratory sounds, wherein:

[0129] Normal: The following conditions are met simultaneously: the respiratory cycle interval is uniform, and the difference between adjacent cycle durations is less than or equal to 0.5 seconds; there is no respiratory airflow interruption exceeding 2 seconds in the entire sample; no high-pitched throat humming sound is heard during inhalation or exhalation; there is no wheezing, snoring, phlegm humming sound, or water bubble sound; the number of coughs within 30 seconds is less than or equal to 1, and the cough is single-voiced and non-stimulating; the average inhalation time accounts for less than 50% of the entire respiratory cycle.

[0130] Low risk: In the entire sample, clear inspiratory throat humming sound can be heard for more than 3 times, but the sound intensity is not high; or 1-3 times of barking or irritating dry cough occurs within 30 seconds. It can be accompanied by changes in respiratory rhythm, with a difference in adjacent respiratory cycle duration greater than 1.5 seconds and a respiratory cycle interval less than 3 seconds; there is no significant decrease in respiratory sound amplitude.

[0131] High risk: In the entire sample, there is at least one respiratory airflow interruption with a duration of more than 5 seconds; the breathing is extremely irregular, with a difference between the longest and shortest cycle durations of more than 2 seconds in 5 consecutive respiratory cycles; the inspiratory throat humming sound is extremely loud and irritating, and lasts for a long time, or there is a biphasic throat humming sound; after a bout of coughing, there is a airflow interruption of more than 2 seconds; compared with the baseline, the respiratory sound amplitude decreases by less than 70% or disappears.

[0132] The labeled respiratory sound dataset is randomly divided into a training set and a test set;

[0133] In the training set, the respiratory sound features are used as input, and the normal, low risk, and high risk are used as output to train the CNN algorithm architecture to obtain a suffocation abnormality classification model;

[0134] As shown in Figure 3 , the CNN6 algorithm architecture includes 4 convolution blocks ConvBlock, 1 feature aggregation layer, and 1 fully connected classification layer, and each convolution block uses a series combination of a convolution layer Conv2D and a maximum pooling layer MaxPooling2D;

[0135] The 4 convolution blocks are used to gradually extract deep features in the respiratory sound features;

[0136] The feature aggregation layer is used to integrate deep features to form a global semantic representation;

[0137] The fully connected classification layer is used to output the classification results of normal, low risk, and high risk.

[0138] The application adopts a CNN6 algorithm framework specially built for audio signal recognition in the AudioSet dataset to construct a classification model. The network is optimized based on the classic convolutional neural network architecture, reduces the computational complexity while retaining high classification accuracy through hierarchical feature extraction and global information fusion, and is suitable for deployment on resource-constrained medical edge devices. The model structure includes 4 convolution blocks (ConvBlock), 1 feature aggregation layer, and 1 fully connected classification layer, as shown in the overall architecture Figure 3 Each convolution block uses a combination of convolution layers (Conv2D) and max-pooling layers (MaxPooling2D) to gradually extract deep features from the audio signal. The size of the convolution kernel is fixed at 3x3, which helps to capture local features while reducing computational complexity. As the convolution blocks deepen, the number of channels increases exponentially (from 64 to 128, 256, and finally to 512), allowing more diverse feature information to be extracted. At the same time, through the pooling operation with a step of 2, the spatial dimension of the feature map is gradually compressed (e.g., from 128x800 to 8x50), which helps to reduce data redundancy and improve feature robustness. In the shallow convolution blocks, the model mainly focuses on local time-frequency texture features of the audio signal, which are crucial for distinguishing different classes of audio signals. As the convolution blocks deepen, the model gradually shifts to capturing energy distribution patterns within a wide time window, which helps to further improve classification accuracy. The high-dimensional features output by the convolution layer often contain redundant information and need to be compressed through pooling. This operation takes the mean or maximum value of the local region, retaining key features while reducing dimensionality, which not only effectively reduces the computational resource consumption of the model, but also suppresses the risk of overfitting caused by data redundancy. The fully connected layer is the core prediction module of the model, which integrates local features extracted by convolution and pooling to form a global semantic representation and finally outputs the classification result. To obtain interpretable class confidence, the Softmax function needs to be applied to the output layer for probability normalization, converting the original score to a probability distribution of each class. During prediction, the class corresponding to the maximum probability value is taken as the final classification result.

[0139] In order to systematically evaluate the performance of the suffocation risk classification model, the following classification performance evaluation indexes are selected for comprehensive analysis: accuracy (Accuracy, Acc), recall (Recall, Rec), precision (Precision, Pre), and F1 score. The four core indexes in the confusion matrix, true positive (True Positive, TP), true negative (True Negative, TN), false positive (False Positive, FP), and false negative (False Negative, FN), are used for model evaluation: true positive (True Positive, TP): the number of real positive samples correctly predicted as positive by the model; true negative (True Negative, TN): the number of real negative samples correctly predicted as negative by the model; false positive (False Positive, FP): the number of real negative samples incorrectly predicted as positive by the model; false negative (False Negative, FN): the number of real positive samples incorrectly predicted as negative by the model;

[0140] The specific calculation methods and clinical significance of each index are as follows:

[0141] Accuracy (Acc): evaluates the overall correct rate of positive and negative predictions of the model for all samples.

[0142] Recall (Rec): evaluates the recognition ability of the model for real positive samples.

[0143] Precision (Pre): reflects the accuracy of the model for predicting positive samples.

[0144] F1 score: a balanced index that measures the accuracy of the model in predicting positive samples, which combines the performance of precision and recall.

[0145] In addition, the receiver operating characteristic curve (Receiver Operating Characteristic, ROC) and its area under the curve (Area Under the Curve, AUC), as well as the precision-recall curve (Precision-Recall Curve, PR), are also used to evaluate the model performance.

[0146] The data set is randomly divided into a training set and a test set according to an 8:2 ratio, a suffocation risk classification model is constructed, and the performance evaluation results of the model on the training set and the test set are obtained, as shown in Figure 4 and 5 .

[0147] like Figure 4 As shown in Figure a, the training loss and testing loss of the model gradually decrease with the increase of training epochs, eventually stabilizing and converging to a low level with an average loss of 0.2174. This indicates that the model effectively learns the data features during training and gradually reduces the prediction error. Figure 4 As shown in Figure b, the training accuracy and test accuracy of the model gradually improve with the increase of training epochs, eventually reaching an overall accuracy of 96.42% on the test set. This indicates that the model exhibits good accuracy and stability in classification tasks.

[0148] From the perspective of classification performance, such as Figure 5 As shown in Figure b, the model performed best in identifying cases without suffocation risk, with a precision of 97.79%, a recall of 99.50%, and an F1 score of 98.65%. This indicates that the model effectively avoids misclassifying healthy children as at-risk, thereby reducing unnecessary medical examinations and parental anxiety. In clinical practice, this helps optimize resource allocation, focusing attention on children who are truly at risk. For data with low suffocation risk, the recall was 97.75%, the precision was 91.78%, and the F1 score was 94.67%. This shows that the model's ability to identify children with low suffocation risk remains strong. Although 8 low-risk cases were misclassified as high-risk, from a safety perspective, such misclassification is more acceptable than missed diagnosis, as it ensures that the risk is not underestimated and is more conducive to the early detection of children with potential risks. For data with high suffocation risk, the precision was 97.61%, the recall was 89.76%, and the F1 score was 93.52%. This indicates that the model has a low false positive rate when identifying high-risk suffocation states; almost all data judged as high-risk suffocation states are high-risk audio, with very few misjudgments. However, from... Figure 5 The confusion matrix of 'a' shows that it misclassified 34 audio segments with high suffocation risk as low suffocation risk, which means that a false negative occurred.

[0149] from Figure 5 The ROC curve of c and Figure 5The PR curve in the middle d can be seen that the model performs best in identifying normal breathing, and the AUC value reaches 0.99, which can almost completely distinguish the normal breathing state from the abnormal breathing state. At the same time, the model has high precision and recall rate, which can accurately identify most normal breathing samples and maintain a low misdiagnosis rate, further verifying its good performance. The model in identifying low asphyxia risk, the AUC value is 0.97, indicating that it has high accuracy. But its precision rate decreases at high recall rate, there is a certain false positive rate, which may lead to misjudgment of asphyxia risk type. The model in identifying high asphyxia risk, the AUC value is 0.94, although it is relatively lower than the above results, but still maintains at a high level, but its ability to identify high asphyxia risk decreases. The above results show that the asphyxia risk classification model constructed by the present application has good overall performance, which can be used to monitor the asphyxia risk of infants. However, in asphyxia risk prediction, sensitivity should be given priority to specificity, because the cost of missed diagnosis is much higher than that of misdiagnosis, so the model can be further optimized to improve its sensitivity to high asphyxia risk state.

[0150] The method for constructing the dynamic baseline model comprises:

[0151] Select a normal breathing sound signal conforming to a specified time length (such as 24 hours) in the denoised breathing sound signal as a learning period signal;

[0152] Obtain the MFCC feature, the fundamental frequency feature, the harmonic energy feature, the expiration time length and amplitude, the inspiration time length and amplitude of the learning period signal, and calculate the mean value and standard deviation of each feature within the specified time length;

[0153] According to the Gaussian distribution of the features within the specified time length, the baseline model of each feature is established , and is combined as a dynamic baseline model , wherein, is the baseline model of the feature z, is the mean value of the feature z, is the standard deviation of the feature z.

[0154] After the dynamic baseline model enters the monitoring state, it will continuously (or at certain time intervals) update with new breathing data judged as normal by the asphyxia risk classification model, such as using the recursive estimation method of exponential smoothing, so that the baseline model slowly adapts to the growth and natural changes of the infant. The exponential smoothing updated mean value is: , wherein, is the updated mean value, is the value of the current feature z, is the mean value before updating, is a very small learning rate, such as 0.001, which ensures that the baseline will not change dramatically due to short-term fluctuations. The dynamic baseline model can also be updated in a similar way. The updated dynamic baseline model is .

[0155] The dynamic baseline model established by the present application is an individual digital portrait of the normal breathing pattern of the infant, which is personalized and continuously self-adapted over time. It is not a fixed threshold, but a statistical model describing the normal fluctuation range of various respiratory sound features of the infant in a healthy state.

[0156] For example, the normal breathing patterns of different infants vary greatly. The respiratory sound of a full-term healthy infant may be loud and powerful, while the respiratory sound of a premature infant or a weaker infant may be naturally very weak. If a uniform and fixed threshold is used (for example, an alarm is given when the audio energy is below a certain value), the system will continuously generate false alarms for infants with weak respiratory sound, and may miss alarms for infants with loud respiratory sound.

[0157] After the dynamic baseline model is constructed, the individualized normal parameter range unique to the current monitored infant is learned and recorded. For example, the normal expiration time baseline of a certain infant is 2s±0.3s, and subsequent monitoring is compared with this infant's own baseline rather than a universal standard, and a warning is given when the baseline level is exceeded to a certain extent, so as to achieve accurate respiratory abnormality warning that adapts to individual characteristics.

[0158] The dynamic baseline model provides personalized monitoring for each infant, recognizes and adapts to physiological diversity, can accompany the growth of the infant, always maintains the accuracy of monitoring, does not need to be manually reset or calibrated, ensures the effectiveness of long-term use, and the dynamic baseline itself is a record of healthy trends. Doctors or parents can view the curve of the baseline over time to understand the respiratory development or recovery process of the infant, rather than just receiving isolated alarm events.

[0159] After the present application detects respiratory abnormalities, the abnormal respiratory sound signal is transmitted to the suffocation risk classification model to distinguish whether the respiratory abnormality belongs to obstructive suffocation or central suffocation, and feedback to the guardian terminal for risk warning.

[0160] The method for monitoring and warning respiratory abnormalities of an infant using a dynamic baseline model and classifying suffocation risks of respiratory abnormalities fed back by the dynamic baseline model using a suffocation abnormality classification model comprises:

[0161] Using the dynamic baseline model, the deviation of the MFCC feature, the fundamental frequency feature, the harmonic energy feature, the expiration time and amplitude, and the inspiration time and amplitude of the real-time denoised respiratory sound signal of the infant is monitored, wherein:

[0162] When the MFCC feature, the fundamental frequency feature, the harmonic energy feature, the expiration time length and amplitude, and the inspiration time length and amplitude of the real-time denoised breathing sound signal of the infant deviate from the dynamic baseline model by more than a preset threshold, it is determined that the infant is in a respiratory abnormal state;

[0163] When the MFCC feature, the fundamental frequency feature, the harmonic energy feature, the expiration time length and amplitude, and the inspiration time length and amplitude of the real-time denoised breathing sound signal of the infant deviate from the dynamic baseline model by not more than a preset threshold, it is determined that the infant is not in a respiratory abnormal state;

[0164] The MFCC feature, the fundamental frequency feature, the harmonic energy feature, the expiration time length and amplitude, and the inspiration time length and amplitude of the denoised breathing sound signal corresponding to the infant in the respiratory abnormal state are input into the asphyxia abnormality classification model to determine the asphyxia risk category of the infant.

[0165] As shown in Figure 2 The present application provides an infant asphyxia risk autonomous monitoring system based on throat breathing sound, which is applied to an infant asphyxia risk autonomous monitoring method based on throat breathing sound, comprising:

[0166] A data acquisition unit comprising a high-sensitivity MEMS microphone and a micro accelerometer, the high-sensitivity MEMS microphone being used to acquire airflow and breathing sound signals of the infant's throat, and the micro accelerometer being used to synchronously monitor thoracic movement data of the infant;

[0167] A signal denoising unit is used to denoise the airflow and breathing sound signals by integrating the airflow and breathing sound signals and the thoracic movement data, so as to obtain denoised breathing sound signals;

[0168] A machine learning unit is used to separate abnormal breathing sound signals from the denoised breathing sound signals, and to combine normal breathing sound signals and abnormal breathing sound signals into a breathing sound data set; breathing sound features are extracted from the breathing sound data set, an asphyxia abnormality classification model distinguishing normal, low-risk and high-risk is constructed based on the breathing sound features through a machine learning model, and a dynamic baseline model adapting to individual respiratory differences is established according to the breathing sound features of the normal breathing sound signals;

[0169] An autonomous monitoring unit is used to monitor and warn the respiratory abnormalities of the infant by using the dynamic baseline model, and to classify the asphyxia risk of the respiratory abnormalities fed back by the dynamic baseline model by using the asphyxia abnormality classification model.

[0170] The present application provides a computer-readable storage medium, and the computer-readable storage medium stores computer execution instructions, when the processor executes the computer execution instructions, the method for autonomous monitoring of infant asphyxia risk based on throat breathing sound is realized.

[0171] The present application directly collects throat breathing sound by using a high-sensitivity MEMS microphone, and reduces environmental noise interference and improves early suffocation identification accuracy through noise reduction processing by a miniature accelerometer;

[0172] The embedded AI algorithm formed by the machine learning model and the dynamic baseline model can effectively distinguish normal breathing variation precursors from suffocation categories, and realize real-time analysis for infants.

[0173] The present application combines the hardware devices of high-sensitivity MEMS microphone and miniature accelerometer with artificial intelligence model, and realizes the self-completion of infant suffocation risk identification and early warning in the home environment by using throat breathing sound signals.

[0174] The above examples are only exemplary embodiments of the present application and are not used to limit the present application, the protection scope of the present application is defined by the claims. Those skilled in the art can make various modifications or equivalent replacements to the present application within the spirit and protection scope of the present application, and such modifications or equivalent replacements are also regarded as falling within the protection scope of the present application.

Claims

1. A method for self-monitoring infant asphyxia risk based on laryngeal breath sounds, characterized in that, Includes the following steps: A high-sensitivity MEMS microphone was used to collect airflow and respiratory sound signals from the infant's throat, and a miniature accelerometer was used to simultaneously monitor the infant's chest movement data. By integrating the airflow and respiratory sound signals with the chest movement data, the airflow and respiratory sound signals are denoised to obtain a denoised respiratory sound signal. The denoised breath sound signals are separated according to breathing events and enhanced and combined into a breath sound dataset; Breath sound features were extracted from the breath sound dataset. A machine learning model was used to construct a classification model for asphyxia anomalies that distinguishes between normal, low-risk, and high-risk individuals based on the breath sound features. A dynamic baseline model adapted to individual breathing differences was also established based on the breath sound features. The dynamic baseline model is used to monitor and warn of respiratory abnormalities in infants, and the asphyxia abnormality classification model is used to classify the asphyxia risk based on the respiratory abnormalities fed back by the dynamic baseline model. A method for noise reduction of the airflow and respiratory sound signals by integrating the airflow and respiratory sound signals with the chest wall movement data includes: The airflow and breath sound signals are decomposed into components, and the airflow and breath sound signals are as follows: ,in, These are airflow and breath sound signals. For the target breath sound signal, for Noise signals generated by body movement, for Environmental noise signals in; The thoracic motion data is decomposed into components, and the thoracic motion data is as follows: ,in, For chest movement data, for Noise signals generated by body movement, for Environmental noise signals in; The adaptive filter output noise is estimated by iterative computation using the Least Mean Square (LMS) algorithm. The target breath sound signal was calculated based on noise estimation. ,in, This is a noise estimate. A collection of chest cavity movement data. , for Chest movement data from the previous moment, for Thoracic motion data for the first L-1 time points, For filter weights, The target breath sound signal; target breath sound signal As the denoised breath sound signal; Methods for constructing dynamic baseline models include: Select a normal breath sound signal that meets the specified duration from the denoised breath sound signal signal as the learning period signal; Obtain the MFCC characteristics, fundamental frequency characteristics, harmonic energy characteristics, expiratory duration and amplitude, and inspiratory duration and amplitude of the signal during the learning period, and calculate the mean and standard deviation of each characteristic within the specified duration; Based on the fact that the features follow a Gaussian distribution over a specified time period, a baseline model is established for each feature. And combined as a dynamic baseline model ,in, The baseline model for feature z. Let z be the mean of the feature z. Let z be the standard deviation of the characteristic z.

2. The method for self-monitoring infant asphyxia risk based on laryngeal breath sounds according to claim 1, characterized in that: The method for separating the denoised breath sound signal based on respiratory events includes: The denoised breathing sound signal is divided into multiple audio samples according to the preset duration, and the audio samples with insufficient preset duration are padded with duration. Calculate the short-time energy (STE) and short-time zero-crossing rate (ZCR) for each audio sample, and eliminate silent audio samples based on the STE and ZCR using a dual-threshold method. Among the remaining audio samples, audio samples corresponding to respiratory events such as inhalation, exhalation, crying, coughing, and abnormal breathing sounds, identified based on short-time energy (STE) and short-time zero-crossing rate (ZCR), are retained as valuable samples.

3. The method for self-monitoring infant asphyxia risk based on laryngeal breath sounds according to claim 2, characterized in that: Methods for enhancing and combining valuable samples into a breath sound dataset include: Time-domain shift: The value sample is shifted backward by 1 second on the time axis, and the 7.0 s to 8.0 s portion of the value sample is spliced ​​to the beginning of the value sample; Gain adjustment: Decrease or increase the amplitude of each value sample, where the oscillation parameter of the amplitude changed by decreasing or increasing is a random value between 0.1 and 5.0; Noise introduction: Add random noise signals to the valuable samples and adjust the signal-to-noise ratio to a preset value; The valuable samples and the enhanced valuable samples are combined to form the respiratory sound dataset.

4. The method for self-monitoring infant asphyxia risk based on laryngeal breath sounds according to claim 3, characterized in that: Methods for extracting breath sound features include: The samples in the breath sound dataset are sequentially pre-emphasized, framed, windowed, subjected to fast Fourier transform, Mel filter bank, and logarithmic energy processing to obtain Mel spectrograms. The Mel spectrograms are then subjected to DCT transform to obtain MFCC features. The fundamental frequency characteristics of the samples in the respiratory sound dataset are estimated using the cepstral method, and the harmonic energy characteristics are calculated based on the fundamental frequency characteristics. The expiratory duration and amplitude, and the inspiratory duration and amplitude are calculated sequentially for the samples in the breath sound dataset; MFCC characteristics, fundamental frequency characteristics, harmonic energy characteristics, expiratory duration and amplitude, and inspiratory duration and amplitude are used as respiratory sound characteristics.

5. The method for self-monitoring infant asphyxia risk based on laryngeal breath sounds according to claim 4, characterized in that: Methods for constructing asphyxiation anomaly classification models include: The samples in the breath sound dataset were labeled as normal, low-risk, and high-risk, and the labeled breath sound dataset was randomly divided into training set and test set. In the training set, the CNN algorithm architecture is trained with respiratory sound features as input and normal, low risk and high risk as output to obtain the asphyxiation anomaly classification model. The CNN6 algorithm architecture includes: 4 convolutional blocks (ConvBlock), 1 feature aggregation layer, and 1 fully connected classification layer. Each convolutional block uses a concatenated combination of a convolutional layer (Conv2D) and a max pooling layer (MaxPooling2D). Four convolutional blocks are used to progressively extract deep features from the breath sound features; The feature aggregation layer is used to integrate deep features to form a global semantic representation; The fully connected classification layer is used to output classification results as normal, low risk, and high risk.

6. The method for self-monitoring infant asphyxia risk based on laryngeal breath sounds according to claim 5, characterized in that: Methods for monitoring and early warning of respiratory abnormalities in infants using dynamic baseline models, and for classifying asphyxia risk using asphyxia anomaly classification models based on respiratory abnormalities reported by dynamic baseline models, include: Using a dynamic baseline model, the deviations of the real-time denoised respiratory sound signal of an infant from its MFCC characteristics, fundamental frequency characteristics, harmonic energy characteristics, expiratory duration and amplitude, and inspiratory duration and amplitude were monitored. If the MFCC characteristics, fundamental frequency characteristics, harmonic energy characteristics, expiratory duration and amplitude, and inspiratory duration and amplitude of the infant's real-time denoised respiratory sound signal deviate from the dynamic baseline model by more than a preset threshold, the infant is determined to be in an abnormal respiratory state. If the MFCC characteristics, fundamental frequency characteristics, harmonic energy characteristics, expiratory duration and amplitude, and inspiratory duration and amplitude of the infant's real-time denoised respiratory sound signal do not deviate from the dynamic baseline model by more than a preset threshold, the infant is determined not to be in an abnormal respiratory state. The MFCC features, fundamental frequency features, harmonic energy features, expiratory duration and amplitude, and inspiratory duration and amplitude of the denoised respiratory sound signal corresponding to the infant's abnormal breathing state are input into the asphyxia anomaly classification model to determine the infant's asphyxia risk category.

7. A self-monitoring system for infant asphyxia risk based on laryngeal breath sounds, characterized in that, The method for self-monitoring infant asphyxia risk based on laryngeal breath sounds, as described in any one of claims 1-6, comprises: The data acquisition unit includes a high-sensitivity MEMS microphone and a miniature accelerometer. The high-sensitivity MEMS microphone is used to collect airflow and respiratory sound signals from the infant's throat, and the miniature accelerometer synchronously monitors the infant's chest movement data. The signal noise reduction unit is used to perform noise reduction processing on the airflow and respiratory sound signals by integrating the airflow and respiratory sound signals and the chest wall movement data to obtain a noise-reduced respiratory sound signal; The machine learning unit is used to separate abnormal breath sound signals from the denoised breath sound signals, and enhance and combine normal breath sound signals and abnormal breath sound signals into a breath sound dataset; extract breath sound features from the breath sound dataset, construct a classification model for asphyxia abnormality based on the breath sound features through a machine learning model to distinguish between normal, low-risk and high-risk, and establish a dynamic baseline model adapted to individual breathing differences based on the breath sound features of the normal breath sound signal; The autonomous monitoring unit is used to monitor and warn of respiratory abnormalities in infants using a dynamic baseline model, and to classify the risk of asphyxiation based on respiratory abnormalities fed back by the dynamic baseline model using an asphyxiation abnormality classification model.

8. The infant asphyxia risk self-monitoring system based on laryngeal breath sounds according to claim 7, characterized in that: The input to the asphyxia anomaly classification model is the breath sound feature, and the output is the asphyxia anomaly category: normal, low risk, and high risk.

Images