A non-contact acoustic detection method for supine body position during overnight sleep

Abstract

This invention discloses a non-contact acoustic detection method for supine sleep position throughout the night. The method first uses a dual-element array of microphones to synchronously collect the sleep sound data of snoring patients throughout the night. Then, it uses sound segment detection and principal component projection (PCP) to automatically detect and segment snoring segments throughout the night. Next, it extracts three spatial features for each snoring segment that are unaffected by head position: the ratio of high-frequency to low-frequency energy between array elements, the deviation of the normalized main peak amplitude of the spatial cross-correlation between array elements, and the deviation of the normalized sidelobe fluctuation. Finally, it performs unsupervised clustering on the spatial feature vector sample set of snoring segments throughout the night. By comparing it with the expected feature vector template, it determines the clustering result cluster corresponding to the head-tilt state, and marks the time corresponding to each snoring segment in the overnight snoring recording data. This invention achieves non-contact acoustic detection of supine sleep position throughout the night by recognizing snoring segments in the head-tilt state.

Description

Technical Field

[0001] This invention relates to the field of biomedical signal analysis and acoustic signal processing, specifically a non-contact acoustic detection method for overnight sleep supine position. Background Technology

[0002] Obstructive sleep apnea / hypopnea syndrome (OSAHS) is a common sleep-disordered breathing condition. Polysomnography (PSG), the gold standard for clinical diagnosis of OSAHS, faces several challenges: First, PSG requires specialized monitoring facilities and expensive equipment, limiting its application to dedicated sleep laboratories in some hospitals. Many smaller hospitals, such as community and town hospitals, lack the necessary resources and expertise to provide timely assessment and widespread screening for a large number of patients. Second, PSG requires patients to wear masks for extended periods and attach dozens of leads to multiple body parts, causing discomfort and impacting overall sleep quality, thus compromising diagnostic accuracy. Therefore, PSG technology is difficult to apply for early screening of potential OSAHS patients, home monitoring of mild to moderate cases, and postoperative follow-up of severe cases. Consequently, there is an urgent need for more economical, convenient, and reliable community screening and home monitoring methods.

[0003] Sleep posture is a crucial factor influencing the severity of obstructive sleep apnea-hypopnea syndrome (OSAHS). The apnea-hypopnea index (AHI) is a key indicator for assessing the severity of OSAHS. Posturally dependent OSAHS is defined as an AHI at least twice that of other sleeping positions when supine. Studies have shown that over 55% of OSAHS patients exhibit postural dependence, and patients with milder symptoms are more likely to have stronger postural dependence. Therefore, supine sleep posture assessment can facilitate more accurate initial screening and sleep intervention for patients with mild to moderate OSAHS exhibiting high postural dependence.

[0004] Among existing non-PSG sleep monitoring methods, non-contact devices, because they do not come into contact with the body and do not cause discomfort to the patient, can obtain medical and physiological information of the patient in a completely natural sleep state. Visual technologies such as visible light and infrared can more intuitively capture the patient's sleep position information and its changes throughout the night, but they pose a risk of leaking the patient's personal and family privacy. Microwave technologies such as radar are relatively difficult to obtain sleep position information and are also controversial regarding electromagnetic radiation hazards. In comparison, acoustic technology can use the directional characteristics of sound emitted from the patient's mouth and nose in space to detect sleep position status. It has advantages such as low cost, simple operation, high privacy and security, making it the preferred solution for achieving real-time, dynamic sleep position detection throughout the night. Since snoring is one of the most common clinical manifestations of OSAHS, and severe habitual snoring is a precursor to OSAHS, non-contact snoring analysis methods have always been a research hotspot in clinical medical snoring detection and electronic information fields.

[0005] Due to the shape of the head and the effects of obstruction and diffraction by other parts of the body, the snoring signals collected by the microphone when facing different sides of the head will exhibit characteristic differences. That is, the snoring signals emitted by snorers through the mouth and nose during sleep are not isotropic in spatial propagation, but rather exhibit certain directional characteristics, with higher frequency snoring signals showing stronger directional propagation. Based on this frequency-dependent spatial propagation directionality of snoring signals, non-contact acoustic detection of sleep position can be achieved by placing one or more microphones on each side of the snorer's bedside in an array. Even if some position-dependent snorers have no snoring or only mild snoring when sleeping on their side, the snoring will reappear or worsen when they turn to their supine position, thus not affecting the acoustic detection of the supine sleeping position.

[0006] The literature (Pu Y, Liu T, Zhao Z, et al. Sleep body position detection based on microphone array[C]. International Conference on Signal Processing and Communication Technology(SPCT 2021),2021:121780Q) discloses a sleep position detection method based on a microphone array. This method uses the high-frequency energy ratio and spatial cross-correlation deviation of the snoring signals between two array elements located on both sides of the bed as signal features for identifying the direction of the sleeping head. Unsupervised clustering proves that the above features have good separability, and synchronous infrared imaging confirms that there is a high degree of consistency between the direction of the sleeping head and the sleep position, thus verifying the feasibility of non-contact acoustic detection of the sleep position of snoring patients. However, this method does not consider the automatic detection of snoring signals throughout the night and the translational changes of the patient's head position during sleep. When the distance between the patient's mouth and nose and the two microphone array elements differs greatly, the method's ability to distinguish different sleep positions will be significantly reduced. Summary of the Invention

[0007] The purpose of this invention is to provide a non-contact acoustic detection method for supine body position during overnight sleep, which can effectively suppress the influence of changes in the patient's head position during the night on the detection results. It is suitable for community screening, home monitoring and postoperative follow-up of obstructive sleep apnea.

[0008] The technical solution to achieve the objective of this invention is as follows: Firstly, this invention provides a non-contact acoustic detection method for supine body position during overnight sleep, comprising the following steps:

[0009] Step 1: Preprocess the sleep sound signals of snoring patients throughout the night collected by a dual-element array consisting of microphones placed on both sides of the bed. Then, use the sound segment detection and principal component projection method of snoring feature values ​​to automatically detect and segment snoring segments throughout the night and establish a data sample set of snoring segments throughout the night.

[0010] Step 2: For the overnight snoring segment data sample set obtained in Step 1, extract three spatial features of each snoring segment that are not affected by head position: the ratio of high and low frequency energy between array elements, the deviation of the normalized main peak amplitude of spatial cross-correlation between array elements, and the deviation of normalized sidelobe fluctuation, to form a three-dimensional spatial feature vector sample set of overnight snoring segments.

[0011] Step 3: Perform unsupervised clustering on the three-dimensional spatial feature vector sample set of snoring fragments extracted in Step 2. Compare the representative samples of the feature vectors in each cluster with the expected feature vector template of the head-tilt state to determine the cluster corresponding to the supine sleeping position. Mark the time corresponding to each snoring fragment in the overnight snoring recording data, thereby realizing non-contact acoustic detection of the supine sleeping position of snoring patients throughout the night.

[0012] In a second aspect, the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method described in the first aspect.

[0013] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in the first aspect.

[0014] Fourthly, the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the method described in the first aspect.

[0015] Compared with existing technologies, the beneficial effects of this invention are as follows: 1) Using a microphone array to collect snoring information from patients with snoring can greatly reduce the "invasive" discomfort for patients. The non-contact signal acquisition process does not affect the operation of medical staff or the patient's sitting, turning, or rolling movements, while also taking into account changes in the patient's position; 2) The nighttime snoring signal data collected by the dual-element microphone array effectively utilizes the spatial propagation directionality of snoring signals, enabling non-contact acoustic detection of supine sleep positions based on multi-dimensional snoring spatial characteristics; 3) The fully automated snoring event detection method based on the collected nighttime sleep sound signal data can effectively... 4) The non-snoring segments in the overnight sleep sound signal data are removed to accurately establish an overnight sleep snoring segment database; 5) The high-low frequency energy ratio between array elements, the deviation of the normalized main peak amplitude and the deviation of the normalized sidelobe fluctuation characteristics of the array elements extracted from the signals of the two array element channels on both sides of the bed can well reflect the spatial propagation directionality of snoring signals caused by the obstruction effect of the head and body and the multipath propagation effect, so as to effectively detect the supine sleeping position, and the detection results are not affected by the position of the sleeping head; 6) The method of the present invention is low in cost, simple to operate, and the monitored object does not feel any "invasive" discomfort.

[0016] The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0017] Figure 1 This is a flowchart of a non-contact acoustic testing method for supine body position during overnight sleep.

[0018] Figure 2 This is a block diagram of the LMS adaptive filtering algorithm.

[0019] Figures 3(a) and 3(b) are schematic diagrams of the coordinates of the patient's head at different positions. Figure 3(a) shows the head position coordinates at the midline of the movement range, and Figure 3(b) shows the head position coordinates at any position within the movement range.

[0020] Figures 4(a) to 4(d) This is an example image showing the detection results of snoring and non-snoring segments in the automatic detection of snoring events throughout the night.

[0021] Figure 5 It is a frequency plot of the high and low frequency energy ratio between array elements in the snoring segment data sample set.

[0022] Figure 6 This is a normalized peak amplitude distribution diagram of the spatial cross-correlation between array elements of a snoring segment data sample set.

[0023] Figure 7 This is a normalized sidelobe fluctuation distribution diagram of the spatial cross-correlation between array elements in a snoring segment data sample set. Detailed Implementation

[0024] Combination Figure 1 The present invention provides a non-contact acoustic detection method for overnight supine sleep position, comprising the following steps:

[0025] Dataset Introduction: Two microphone arrays are placed on either side of the bedside, with a signal sampling rate of 16kHz and a quantization precision of 16bit. A simulated snoring segment data sample set is established using overnight sleep recordings of a middle-aged female patient with mild OSAHS as an example. First, based on synchronously recorded infrared images, 200 single-channel snoring segments with a duration of at least 10 minutes in a stable supine sleeping position are extracted, forming the supine position channel one snoring segment data sample set D1. D1 is copied to form the supine position channel two snoring segment data sample set D2. D1 and D2 are combined to form the supine position dual-channel snoring segment data sample set D. y 200 pairs of dual-channel snoring segments were simulated when the patient's head was located on the central axis of the movement range shown in Figure 3(a) in a stable supine sleeping position. Then, D1 was used as the data sample set D3 of the snoring segment of channel one (mouth and nose facing the array element) in the lateral position. A low-pass filter was designed to filter D3 according to the spatial propagation directionality of the snoring signal to form the data sample set D4 of the snoring segment of channel two (mouth and nose facing away from the array element) in the lateral position. D3 and D4 were combined to form the data sample set D of the dual-channel snoring segment in the lateral position. cThe simulation included 200 pairs of dual-channel snoring segments when the patient's head was positioned at the midline of the range of movement shown in Figure 3(a) in a stable lateral sleeping position. Finally, for supine (lateral) sleeping positions, a random coordinate L was generated based on the range of head movement shown in Figure 3(b). i From D y (D c A pair of dual-channel snoring data samples were randomly selected from the data. Based on the influence of head position coordinates on the amplitude and time delay of the snoring signals received by the array elements placed on both sides of the bed, amplitude and time delay corrections were performed on the dual-channel snoring signals to obtain the position of the patient's head at coordinate L when lying supine (or lateral). i A pair of dual-channel snoring sound segment data samples were obtained. The final simulation yielded a dual-channel snoring sound segment data sample set with 500 random head position coordinates in each of the three sleep positions: supine, left lateral, and right lateral.

[0026] Step 1: Preprocess the sleep sound signals of snoring patients collected synchronously by a dual-element array consisting of microphones placed on both sides of the bed. Then, use the sound segment detection and principal component projection method of snoring feature values ​​to automatically detect and segment snoring segments throughout the night, and establish a data sample set of snoring segments throughout the night. The specific steps include:

[0027] Step 1-1: Pre-emphasis, to compensate for the suppression of high-frequency components by the sound generation and sound propagation system, while reducing low-frequency interference.

[0028] Steps 1-2: Framing and Windowing. The q-th frame signal obtained after framing the time-domain signal x(n) of the entire night's recording data is represented as follows:

[0029] x q (n)=w(n)x((q-1)L s +n), 1≤n≤N, 1≤q≤Q x

[0030] In the formula, w(n) is the window function, x q (n) is the value of the nth data sample in the qth frame, where n = 1, 2, ..., N, and q = 1, 2, ..., Q. x N is the number of samples corresponding to the frame length, Q x L represents the total number of frames after the time-domain signal of the entire night's recording data is divided into frames. s For the number of samples corresponding to the frame shift, this invention takes N = 1024, corresponding to 64ms, L s =512, corresponding to 32ms.

[0031] Steps 1-3: Calculate the short-time energy of the q-th frame signal.

[0032]

[0033] Then, two threshold values, short-time energy and consecutive frame count (i.e. signal width), are used to detect audio segments in the overnight recording data, resulting in a sample set of audio segments from the entire night's sleep.

[0034] Steps 1-4: Use the principal component projection method based on snoring feature values ​​to remove non-snoring data frames from the audible segments detected in Steps 1-3. First, randomly extract multiple snoring data segments relatively evenly from the entire night's recording by manual listening. For each snoring data segment y(n), create a Hankel matrix Y.

[0035]

[0036] definition Q y y represents the total number of frames after framing the time-domain signal of this snoring data segment. i =[y(i),y(i+1),…,y(i+N-1)] T Let be the snapshot vector of the i-th sample of the snoring sound. Let Obtain the normalized snoring data matrix

[0037]

[0038] Next, the covariance matrix R of the snoring signal is calculated for all extracted snoring data segments.

[0039]

[0040] Here, avg(·) represents the averaging of all snoring data segments extracted from the entire night's recording. Eigenvalue decomposition is performed on the covariance matrix R of the snoring signal.

[0041] R=UΣU T

[0042] Σ is a diagonal matrix composed of all eigenvalues, and U is the eigenvector matrix composed of the eigenvalues. The eigenvalues ​​are arranged in descending order, and the eigenvectors (i.e., principal components) corresponding to the k largest eigenvalues ​​are used to construct the principal component subspace U of the snoring signal. a In this invention, k = 30.

[0043] Next, the temporal signal z(n) of each audio segment in the overnight recording data obtained in steps 1-3 is processed into frames, and the data matrix Z is constructed according to the following method.

[0044]

[0045] definition Q z z is the total number of frames after the time-domain signal of the audio segment is framed. i=[z((i-1)N / 2+1),z((i-1)N / 2+2),…,z((i+1)N / 2)] T Let be the snapshot vector of the i-th sample in the audio segment. Calculate the projection of the temporal signal of each frame of the audio segment onto the principal component subspace of the snoring signal.

[0046]

[0047] By performing dual-threshold detection of projection energy and projection width on the projection results of the time-domain signal frames of all sound segments in the overnight recording data onto the principal component subspace of the snoring signal, a sample set of overnight sleep snoring segment data is obtained.

[0048] Figure 4(a) shows the principal component projection detection results of snoring feature values ​​from overnight sleep sound data. Figure 4(a) shows the dual-threshold detection results of snoring feature value principal component projection for a segment of snoring sound. Figures 4(b), 4(c), and 4(d) show the dual-threshold detection results of snoring feature value principal component projection for three different non-snoring sound segments: turning over sound, dry cough sound, and walking sound, respectively. Analysis of the detection results shows that the snoring feature value principal component projection amplitude of snoring sound segments is generally high, with an average of around 0.5, and the projection amplitude of waveform frames within the entire snoring sound segment is relatively uniform. In contrast, the snoring feature value principal component projection amplitude of most waveform frames in other non-snoring sound segments is lower than that of snoring sound segments, with an average of only around 0.2. Occasionally, waveform frames with high projection amplitude appear, but the number of consecutive waveform frames is small, failing to meet the set minimum number of frames passing the threshold. This method can effectively remove non-snoring sound segments from overnight sleep sound data, achieving the goal of automatic detection of overnight snoring events.

[0049] Steps 1-5: For each snoring segment detected in Steps 1-4, extract a short, continuous L segment with the highest energy. r The waveform sample vector of this snoring sound segment is composed of individual sample points.

[0050]

[0051] In the formula, the subscript j represents the element index, the superscript T represents the vector transpose, and L r The relationship between N and L satisfies L r =3N; for x j Perform a short-time Fourier transform to obtain the spectrum X of the snoring sound segment. j The present invention takes L r =3072, corresponding to 192ms.

[0052] Step 2: For the overnight snoring segment data sample set obtained in Step 1, extract three spatial features of each snoring segment that are not affected by head position: the ratio of high-frequency to low-frequency energy between array elements, the deviation of the normalized main peak amplitude of the spatial cross-correlation between array elements, and the deviation of the normalized sidelobe fluctuation. This constitutes a three-dimensional spatial feature vector sample set of overnight snoring segments, specifically including the following steps:

[0053] Step 2-1: For each snoring segment, extract the high-low frequency energy ratio characteristics between the two microphone array elements located on both sides of the bed.

[0054]

[0055] Among them, P 1h P 2h P represents the high-frequency energy of the snoring signal received by the two array elements. 1l P 2l The low-frequency energy of the snoring signal received by the two array elements

[0056]

[0057]

[0058]

[0059]

[0060] Among them, X j,k Let k be the value of the k-th frequency point of the snoring spectrum of the j-th array element. h,max k h,min To calculate the frequency index corresponding to the upper and lower bounds of the high-frequency energy range, snoring signals within this frequency band exhibit a clear spatial directionality; k l,max k l,min To calculate the frequency index corresponding to the upper and lower bounds of the low-frequency energy range, snoring signals in this frequency band have no spatial directionality. Because P 1h P 2h P 1l P 2l Both are inversely proportional to the square of the signal propagation distance, and the energy ratio of high and low frequencies, P... 1h / P 1l P 2h / P 2l P will be eliminated respectively 1h P 2h The propagation distance is an influencing factor. It can be seen that the characteristics... This invention can eliminate the influence of head position translation on the high-frequency energy ratio of snoring signals between pairs, while preserving the influence of propagation directionality. The present invention uses k... h,max =208, corresponding to 6500Hz, kh,min =128, corresponding to 4000Hz, k l,max =10, corresponding to 300Hz, k l,min =3, corresponding to 100Hz.

[0061] Figure 5 The frequency plot of the high-low frequency energy ratio between array elements in the snoring segment data sample set is presented, with the ratios rounded to their reciprocals. It can be observed that the ratios are concentrated around 1 in the supine position, with a high degree of clustering, while in the non-supine position, the ratios are concentrated around 0, with a lower degree of clustering compared to the supine position. The main peaks of the two frequency curves in the figure do not overlap, indicating that the high-low frequency energy ratio between array elements has good detection capability in the supine position.

[0062] Step 2-2: Using the two microphone array elements placed on both sides of the bed as reference array elements, calculate the spatial cross-correlation function between array elements using the LMS adaptive filtering method. Then, perform low-pass filtering on the signals of the two array elements and calculate the low-frequency spatial cross-correlation function between array elements using the LMS adaptive filtering method. Finally, extract the normalized main peak amplitude deviation feature and the normalized sidelobe fluctuation deviation feature of the spatial cross-correlation between array elements for each snoring segment.

[0063] Step 2-2-1: Construct a snapshot vector for each snoring segment.

[0064] x j,n =[x j (n),x j (n+1),…,x j (n+L-1)] T j = 1, 2

[0065] In the formula, the subscript j represents the array element number, and n = 1, 2, ..., L r -L+1 represents the snapshot number, where L < L r Let Q be the dimension of the snapshot vector, which is also the length of the adaptive filter; let Q = L be the number of snapshots. r -L+1, in order to maintain the stability of the adaptive filter, Q>3L must be satisfied. In this invention, L=513 is chosen.

[0066] For each snoring segment, use the following method: Figure 2 The LMS adaptive filtering method shown calculates the spatial cross-correlation function between array elements.

[0067]

[0068] in

[0069]

[0070]

[0071] Spatial cross-correlation function In practice, the linear filter coefficient vector of the j1st array element is used to estimate the j2nd array element signal. Due to the occlusion effect of the head and body and the multipath propagation effect, the snoring signal spectrum between the array elements in front of and behind the head is significantly different when sleeping in a side-lying position, especially in the high-frequency part, which leads to w 12 ≠w 21 However, when sleeping in a supine position, the snoring signal can travel directly from the mouth and nose to the microphones on both sides of the bed. 12 and w 21 The differences between them have decreased significantly;

[0072] Step 2-2-2: Low-pass filter and time delay correction are applied to the two array element signals. For each snoring segment, the LMS adaptive filtering method in Step 2-2-1 is used to calculate the low-frequency spatial cross-correlation function w between array elements. 12l and w 21l ;

[0073] Step 2-2-3: For the two microphone array elements placed on either side of the bed headboard, extract the full-band spatial cross-correlation function w between the array elements. 12 and w 21 main peak amplitude p 12 and p 21

[0074]

[0075] Step 2-2-4: For the two microphone array elements placed on either side of the bed headboard, extract the full-band spatial cross-correlation function w between the array elements. 12 and w 21 Sidelobe fluctuations 12 and s 21

[0076]

[0077] in, It is by The sidelobe vector of the full-band spatial cross-correlation function obtained by setting the main peak and one point before and after it to zero;

[0078] Step 2-2-5: For the two microphone array elements placed on either side of the bed headboard, extract the low-frequency spatial cross-correlation function w between the array elements. 12l and w 21l main peak amplitude p 12l and p 21l

[0079]

[0080] Step 2-2-6: For the two microphone array elements placed on either side of the bed, extract the low-frequency spatial cross-correlation function w between the array elements. 12l and w 21l Sidelobe fluctuations 12l and s 21l

[0081]

[0082] in, It is by The sidelobe vector of the low-frequency spatial cross-correlation function obtained by setting the main peak and one point before and after it to zero;

[0083] Step 2-2-7: For the two microphone array elements placed on either side of the bed, calculate the spatial cross-correlation normalized main peak amplitude p. 12 / p 12l and p 21 / p 21l The deviation angle relative to the diagonal in its two-dimensional parameter space yields the normalized main peak amplitude deviation characteristics of the spatial cross-correlation between array elements.

[0084] P ang =|45°-atan((p 12 / p 12l ) / (p 21 / p 21l ))|

[0085] Computational spatial cross-correlation normalized sidelobe oscillations s 12 / s 12l and s 21 / s 21l The deviation angle relative to the diagonal in its two-dimensional parameter space yields the normalized sidelobe fluctuation deviation characteristics of spatial cross-correlation between array elements.

[0086] S ang =|45°-atan((s) 12 / s 12l ) / (s 21 / s 21l ))|

[0087] Since the intensity of snoring signals at all frequencies is uniformly affected by propagation distance, but high-frequency snoring signals exhibit a clear spatial directionality while low-frequency snoring signals lack this directionality, the amplitude of the normalized main peak of the spatial cross-correlation between array elements deviates from the characteristic P. ang and normalized sidelobe oscillation deviation characteristic S ang It can eliminate the effects of head position translation;

[0088] Figure 6 and Figure 7The normalized main peak amplitude p of the spatial cross-correlation between array elements in the snoring segment data sample set is given respectively. 12 / p 12l p 21 / p 21l and normalized sidelobe fluctuations s 12 / s 12l s 21 / s 21l A scatter plot in two-dimensional space and its deviation angle P relative to the diagonal in two-dimensional parameter space. ang S ang Frequency plot. Analyzing the two-dimensional scatter plot reveals the two-dimensional fluctuation points (p) of the normalized main peak amplitude in the supine position. 21 / p 21l ,p 12 / p 12l ) and normalized sidelobe oscillation two-dimensional oscillation point (s 21 / s 21l ,s 12 / s 12l The peak amplitudes of the normalized main peaks are all distributed along the diagonal, rather than in the supine position, and are two-dimensional fluctuation points (p). 21 / p 21l ,p 12 / p 12l ) and normalized sidelobe oscillation two-dimensional oscillation point (s 21 / s 21l ,s 12 / s 12l The snoring signals are distributed diagonally to the upper left and lower right, concentrated near two symmetrical angles. The main reason is that when sleeping on your back, the snoring signal can directly reach the microphones on both sides of the bed from the mouth and nose during propagation. Therefore, the choice of reference array elements has little impact on the spatial cross-correlation function obtained by adaptive filtering in different directions. However, when sleeping on your left side, the snoring signal can directly reach microphone 2, but during propagation to microphone 1, it is significantly affected by head and body obstruction effects and multipath propagation effects. This results in a large difference in the time-frequency characteristics between the snoring signals received by the two microphones, leading to significant differences in the spatial cross-correlation functions obtained by adaptive filtering in different directions using different reference array elements, and correspondingly, greater differences in the two-dimensional deviation characteristics. Analyzing the corresponding frequency diagrams reveals that the P values ​​for supine and non-supine positions... ang Frequency curve and S ang The fact that the main peaks of the frequency curves do not overlap indicates that the deviation of the normalized main peak amplitude and the deviation of the normalized sidelobe fluctuations in the spatial cross-correlation between array elements have a good ability to detect supine body position.

[0089] Step 2-2-8: For the overnight snoring segment data sample set obtained in Step 1, the ratio of high and low frequency energy between array elements, the deviation of the normalized main peak amplitude of the spatial cross-correlation between array elements, and the deviation of the normalized sidelobe fluctuation are used as the three-dimensional spatial feature vector samples of each snoring segment. The three-dimensional spatial feature vector samples of all snoring segments constitute the three-dimensional spatial feature vector sample set of the overnight snoring segments.

[0090] Step 3: Perform unsupervised clustering on the three-dimensional spatial feature vector sample set of snoring fragments extracted in Step 2. Compare the representative samples of the feature vectors in each cluster with the expected feature vector template of the head-tilt state to determine the clusters corresponding to the supine sleep position. Mark the time corresponding to each snoring fragment in the overnight snoring recording data, thereby realizing non-contact acoustic detection of the supine sleep position of snoring patients throughout the night. The specific steps include:

[0091] Step 3-1: For the three-dimensional spatial feature vector sample set C of the snoring fragments extracted in Step 2, unsupervised clustering is performed using the K-means algorithm. The number of cluster centers K is set. Since the samples only contain two different sleep positions that need to be distinguished, supine and non-supine, this invention takes K=2. That is, the feature sample set C is divided into only 2 clusters by clustering, denoted as C={C1,C2}. Based on the prior knowledge of the spatial features of the head-tilt sleep state, the expected feature vector template of the head-tilt sleep state is obtained. The feature vector representative samples in each cluster are compared with the expected feature vector template of the head-tilt sleep state to determine the cluster corresponding to the supine sleep position.

[0092] The clustering results are shown in Table 1. The 500 samples in cluster C1 correspond to supine positions, while the 1000 samples in cluster C2 correspond to non-supine positions. Although theoretically, microphone array technology can only detect head orientation, synchronous infrared monitoring images confirmed a high degree of consistency between the detection results and sleep position orientation. Furthermore, the spatial characteristics unaffected by head position effectively suppress the head translational effects when the patient is in different sleep positions, significantly improving the accuracy of supine position detection. These results demonstrate that the sleep position acoustic features proposed in this invention can be used for non-contact acoustic detection of the sleep position of snoring patients throughout the night.

[0093] Table 1. K-means clustering results based on acoustic features of sleep position

[0094]

[0095] Step 3-2: For the overnight snoring segment data sample set obtained in Step 1, based on the corresponding time of each snoring segment in the overnight recording data, determine the corresponding time of each snoring segment in the cluster of the corresponding supine sleep position obtained in Step 3-1, and mark them respectively in the overnight snoring recording data. If the duration of the marked continuous snoring segments is less than a preset threshold, then the marking of these snoring segments is canceled. Based on the distribution range of the corresponding times of the marked remaining snoring segments, the detection results of the supine sleep period of the snoring patient during the entire night's sleep are obtained.

[0096] In summary, this invention provides a non-contact acoustic detection method for overnight supine sleep position. In this invention, a dual-element array consisting of microphones placed on both sides of the bed can non-contactly collect overnight sleep sound data from patients with snoring. The extracted spatial features—the high-low frequency energy ratio between array elements, the deviation of the normalized main peak amplitude of the spatial cross-correlation between array elements, and the deviation of the normalized sidelobe fluctuation—can effectively reflect the spatial propagation directionality of snoring signals caused by the occlusion effect and multipath propagation effect of the head and body. Therefore, it can effectively detect supine sleep position, and the detection results are not affected by the head position during sleep. Simulation results show that the head orientation obtained by unsupervised clustering is highly consistent with the body orientation displayed by infrared images. Therefore, this invention achieves non-contact acoustic detection of supine sleep position throughout the night by identifying snoring segments in the head-tilt state. This invention is easy to promote, and the proposed non-contact acoustic detection method for overnight supine sleep position has high universality and is of great significance for further realizing economical and rapid community screening, home monitoring, and postoperative follow-up of obstructive sleep apnea.

Claims

1. A non-contact acoustic detection method for supine body position during overnight sleep, characterized in that, Includes the following steps: Step 1: Preprocess the sleep sound signals of snoring patients throughout the night collected by a dual-element array consisting of microphones placed on both sides of the bed. Then, use the sound segment detection and principal component projection method of snoring feature values ​​to automatically detect and segment snoring segments throughout the night and establish a data sample set of snoring segments throughout the night. Step 2: For the overnight snoring segment data sample set obtained in Step 1, extract three spatial features of each snoring segment that are not affected by head position: the ratio of high and low frequency energy between array elements, the deviation of the normalized main peak amplitude of spatial cross-correlation between array elements, and the deviation of normalized sidelobe fluctuation, to form a three-dimensional spatial feature vector sample set of overnight snoring segments. Step 3: Perform unsupervised clustering on the three-dimensional spatial feature vector sample set of snoring fragments extracted in Step 2. Compare the representative samples of the feature vectors in each cluster with the expected feature vector template of the head-tilt state to determine the cluster corresponding to the supine sleeping position. Mark the time corresponding to each snoring fragment in the overnight snoring recording data, thereby realizing non-contact acoustic detection of the supine sleeping position of snoring patients throughout the night.

2. The non-contact acoustic detection method for supine body position during overnight sleep according to claim 1, characterized in that, Step 1 involves preprocessing the sleep sound signals of snoring patients collected by a dual-element array consisting of microphones placed on both sides of the bed. Then, the sound segment detection and principal component projection method of snoring feature values ​​are used to automatically detect and segment snoring segments throughout the night, establishing a data sample set of snoring segments throughout the night. The specific steps include: Step 1-1: Pre-emphasis to compensate for the suppression of high-frequency components by the sound generation and propagation system; Steps 1-2: Framing and Windowing. The q-th frame signal obtained after framing the time-domain signal x(n) of the entire night's recording data is represented as x. q (n)=w(n)x((q-1)L s +n), 1≤n≤N, 1≤q≤Q x In the formula, w(n) is the window function, x q (n) is the value of the nth data sample in the qth frame, where n = 1, 2, ..., N, and q = 1, 2, ..., Q. x N is the number of samples corresponding to the frame length, Q x L represents the total number of frames after the time-domain signal of the entire night's recording data is divided into frames. s The number of samples corresponding to the frame shift; Steps 1-3: Calculate the short-time energy of the q-th frame signal. Then, two threshold values, short-time energy and consecutive frame count, are used to detect audio segments in the overnight recording data to obtain a sample set of audio segments of sleep data throughout the night. Steps 1-4: Use the principal component projection method based on snoring feature values ​​to remove non-snoring data frames from the audible segments detected in Steps 1-3. First, randomly extract multiple snoring data segments relatively evenly from the entire night's recording by manual listening. For each snoring data segment y(n), create a Hankel matrix Y. Define Y = [y1, y2, ..., y Qy ], where Q y y represents the total number of frames after framing the time-domain signal of this snoring data segment. i =[y(i),y(i+1),…,y(i+N-1)] T Let be the snapshot vector of the i-th sample of the snoring sound; let Obtain the normalized snoring data matrix Next, the covariance matrix R of the snoring signal is calculated for all extracted snoring data segments. Where avg(·) represents the averaging of all snoring data segments extracted from the entire night's recording; and eigenvalue decomposition is performed on the covariance matrix R of the snoring signal. R=UΣU T Σ is a diagonal matrix composed of all eigenvalues, and U is the eigenvector matrix composed of the eigenvectors corresponding to each eigenvalue. The eigenvalues ​​are arranged in descending order, and the eigenvectors corresponding to the k largest eigenvalues ​​are used to construct the principal component subspace U of the snoring signal. a ; Next, the temporal signal z(n) of each audio segment in the overnight recording data obtained in steps 1-3 is processed into frames, and the data matrix Z is constructed according to the following method. Define Z = [z1, z2, ..., z Qz ], where Q z z is the total number of frames after the time-domain signal of the audio segment is framed. i =[z((i-1)N / 2+1),z((i-1)N / 2+2),…,z((i+1)N / 2)] T Let i be the snapshot vector of the i-th sample in the audio segment; calculate the projection of the temporal signal of each frame of the audio segment onto the principal subspace of the snoring signal. The projection results of the time-domain signal frames of all sound segments in the overnight recording data in the snoring signal principal cell subspace are subjected to dual threshold detection of projection energy and projection width to obtain a sample set of overnight sleep snoring segment data. Steps 1-5: For each snoring segment detected in Steps 1-4, extract a short, continuous L segment with the highest energy. r The waveform sample vector of this snoring sound segment is composed of individual sample points. x j =[x j,1 ,x j,2 ,…,x j,Lr ] T ,j=1,2 In the formula, the subscript j represents the element index, the superscript T represents the vector transpose, and L r The relationship between N and L satisfies L r =3N; for x j Perform a short-time Fourier transform to obtain the spectrum X of the snoring sound segment. j .

3. The non-contact acoustic detection method for supine body position during overnight sleep according to claim 2, characterized in that, Step 2 involves extracting three spatial features from the overnight snoring data sample set obtained in Step 1, unaffected by head position: the ratio of high-frequency to low-frequency energy between array elements, the deviation of the normalized main peak amplitude of the spatial cross-correlation between array elements, and the deviation of the normalized sidelobe fluctuation. These features form a three-dimensional spatial feature vector sample set for the overnight snoring segments. The specific steps include: Step 2-1: For each snoring segment, extract the high-low frequency energy ratio characteristics between the two microphone array elements located on both sides of the bed. Among them, P 1h P 2h P represents the high-frequency energy of the snoring signal received by the two array elements. 1l P 2l The low-frequency energy of the snoring signal received by the two array elements; Among them, X j,k Let k be the value of the k-th frequency point of the snoring spectrum of the j-th array element. h,max k h,min The frequency index corresponding to the upper and lower bounds of the frequency range for calculating high-frequency energy; k l,max k l,min To calculate the frequency index corresponding to the upper and lower bounds of the low-frequency energy range, snoring signals in this frequency band have no spatial directionality; because P 1h P 2h P 1l P 2l Both are inversely proportional to the square of the signal propagation distance, and the energy ratio of high and low frequencies, P... 1h / P 1l P 2h / P 2l P will be eliminated respectively 1h P 2h The propagation distance influencing factor; Step 2-2: Take the two microphone array elements on both sides of the bed as reference array elements, use the LMS adaptive filtering method to calculate the spatial cross-correlation function between array elements, then perform low-pass filtering on the two array element signals, use the LMS adaptive filtering method to calculate the low-frequency spatial cross-correlation function between array elements, and then extract the normalized main peak amplitude deviation feature and normalized sidelobe fluctuation deviation feature of the spatial cross-correlation between array elements for each snoring segment. Step 2-2-1: Construct a snapshot vector for each snoring segment. x j,n [x j ( n ), x j (n+1),…,x j (n+L-1)] T ,j=1,2 In the formula, the subscript j represents the array element number, and n = 1, 2, ..., L r -L+1 represents the snapshot number, where L < L r Let Q be the dimension of the snapshot vector, which is also the length of the adaptive filter; let Q = L be the number of snapshots. r -L+1, satisfying Q>3L; For each snoring segment, the spatial cross-correlation function between array elements is calculated using the LMS adaptive filtering method. in Spatial cross-correlation function In fact, the linear filter coefficient vector of the j1st array element is used to estimate the signal of the j2nd array element. Step 2-2-2: Low-pass filter the two array element signals and perform time delay correction. For each snoring segment, use the LMS adaptive filtering method from Step 2-2-1 to calculate the low-frequency spatial cross-correlation function w between array elements. 12l and w 21l ; Step 2-2-3: For the two microphone array elements placed on either side of the bed, extract the full-band spatial cross-correlation function w between the array elements. 12 and w 21 main peak amplitude p 12 and p 21 Step 2-2-4: For the two microphone array elements placed on either side of the bed headboard, extract the full-band spatial cross-correlation function w between the array elements. 12 and w 21 Sidelobe fluctuations 12 and s 21 in, It is by The sidelobe vector of the full-band spatial cross-correlation function obtained by setting the main peak and one point before and after it to zero; Step 2-2-5: For the two microphone array elements placed on either side of the bed headboard, extract the low-frequency spatial cross-correlation function w between the array elements. 12l and w 21l main peak amplitude p 12l and p 21l Step 2-2-6: For the two microphone array elements placed on either side of the bed, extract the low-frequency spatial cross-correlation function w between the array elements. 12l and w 21l Sidelobe fluctuations 12l and s 21l in, It is by The sidelobe vector of the low-frequency spatial cross-correlation function obtained by setting the main peak and one point before and after it to zero; Step 2-2-7: For the two microphone array elements placed on either side of the bed, calculate the spatial cross-correlation normalized main peak amplitude p. 12 / p 12l and p 21 / p 21l The deviation angle relative to the diagonal in its two-dimensional parameter space yields the normalized main peak amplitude deviation characteristics of the spatial cross-correlation between array elements. P ang =|45°-atan((p 12 / p 12l ) / (p 21 / p 21l ))| Computational spatial cross-correlation normalized sidelobe oscillations s 12 / s 12l and s 21 / s 21l The deviation angle relative to the diagonal in its two-dimensional parameter space yields the normalized sidelobe fluctuation deviation characteristics of spatial cross-correlation between array elements. S ang =|45°-atan((s 12 / s 12l ) / (s 21 / s 21l ))| Step 2-2-8: For the overnight snoring segment data sample set obtained in Step 1, the high-low frequency energy ratio between array elements, the deviation of the normalized main peak amplitude of the spatial cross-correlation between array elements, and the deviation of the normalized sidelobe fluctuation of each snoring segment are taken as the three-dimensional spatial feature vector samples of that snoring segment; the three-dimensional spatial feature vector samples of all snoring segments constitute the three-dimensional spatial feature vector sample set of the overnight snoring segments.

4. The non-contact acoustic detection method for supine body position during overnight sleep according to claim 3, characterized in that, Step 3 involves unsupervised clustering of the three-dimensional spatial feature vector sample set of snoring fragments extracted in Step 2. The representative samples of the feature vectors in each cluster are compared with the expected feature vector template for the head-tilt position to determine the cluster corresponding to the supine sleep position. The time corresponding to each snoring fragment in this cluster is marked in the overnight snoring recording data, thus achieving non-contact acoustic detection of the supine sleep position of snoring patients throughout the night. This specifically includes the following steps: Step 3-1: For the three-dimensional spatial feature vector sample set C of the snoring fragments extracted in Step 2, unsupervised clustering is performed using the K-means algorithm. The number of cluster centers K is set. Based on the prior knowledge of the spatial features of the head-tilt sleep state, the expected feature vector template of the head-tilt state is obtained. The feature vector representative samples in each cluster are compared with the expected feature vector template of the head-tilt state to determine the cluster corresponding to the supine sleep position. Step 3-2: For the overnight snoring segment data sample set obtained in Step 1, determine the corresponding time of each snoring segment in the cluster of corresponding supine sleep positions obtained in Step 3-1 according to the corresponding time of each snoring segment in the overnight recording data, and mark them in the overnight snoring recording data respectively; if the duration of the marked continuous snoring segments is less than the preset threshold, then the marking of these snoring segments is canceled; based on the distribution range of the corresponding time of the marked remaining snoring segments, obtain the detection results of the supine sleep period of the snoring patient during the whole night's sleep.

5. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method as described in any one of claims 1-4.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method as described in any one of claims 1-4.

7. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the method described in any one of claims 1-4.

Images