Sleep apnea intelligent identification method based on hybrid neural network

Abstract

The invention belongs to the technical field of sleep respiratory disease detection, and relates to a sleep apnea identification method based on a hybrid neural network, which comprises the following steps of: sensing human sleep in a bed surrounding environment by using a millimeter wave radar, and positioning a thoracic cavity position by adopting spatial two-dimensional analysis; the problem that interference exists in radar sign signal extraction of different targets at the same distance is solved. And a respiration signal Shannon entropy is constructed as an iterative index of the whale optimization algorithm, and the index is minimized to realize iterative optimization of variational mode extraction parameters, so that different respiration states are matched with different parameters, and thus a dominant frequency narrowband respiration signal is accurately extracted. A breathing time-frequency energy diagram obtained by data processing is used as input, and a CNN network and a bidirectional long-short-term memory module are used together to construct a feature extraction module, so that the method can recognize the change trend of frequency along with time, capture more comprehensive features, and recognize target sleep apnea more accurately. In addition, the hybrid network enables accurate classification of unmanned scenarios, apnea, hypopnea, and normal respiration.

Description

Technical Field

[0001] The present invention relates to the technical field of non-contact radar sleep monitoring, and in particular to an intelligent monitoring method for sleep apnea of ​​a patient with respiratory disorder at home based on a frequency modulated continuous wave radar. Background Art

[0002] Good quality sleep is essential for health and well-being. However, many disorders during sleep are fatal, such as cardiac arrest, sleep apnea, and asthma. A recent study also showed that breathing disorders are an important cause of sudden infant death syndrome (SIDS) during sleep. Obstructive sleep apnea (OSA) is the most common type of sleep disorder, consisting of multiple episodes of partial or complete breathing obstruction (hypopnea, apnea) during sleep and undiagnosed problems. In many cases, patients with sleep apnea only experience symptoms for a short period of time or on occasional occasions. For patients with this type of disease, long-term hospitalization and observation is undoubtedly unrealistic. Therefore, continuous and cost-effective apnea monitoring in the home sleep environment is essential.

[0003] Traditional solutions mostly use polysomnography, electrocardiogram, etc. for judgment and identification. However, the acquisition of such signals requires a large number of electrodes to be attached to the subject, which is not suitable for home environments. Solutions based on pressure or acceleration sensors require contact with the human body, while lighting conditions limit solutions based on computer vision. Because there is no reliable portable home sleep monitoring system, the Doppler radar system as a home sleep monitoring system has received increasing attention due to its non-contact and unobtrusive monitoring form.

[0004] The use of millimeter-wave radar to obtain respiratory signals for sleep apnea recognition has the advantage of being able to achieve precise positioning with ultra-high position resolution without using imaging technology and leaking user privacy, easily detecting the location of the target and capturing information about the target. In the acquisition and processing of radar target phase signals, the distance range of the target is directly selected in the distance domain information through fast time axis distance FFT, and subsequent signal processing is performed on the phase information of the distance range. However, this method of locating the target only from the distance not only extracts more environmental clutter in the signal, but also causes inaccurate positioning of the area where the human chest vibration is located. When there are multiple human chest targets in the same range gate, the radar signals of different targets are mixed, which affects the apnea detection of a single target. In terms of the input of characteristic signals for identification and detection, compared with heartbeat and heart rate variability (HRV), respiratory signals are rarely used due to their large individual differences and state fluctuations. Inaccurate extraction of respiratory signals can easily lead to apnea recognition errors in some special breathing pattern groups, resulting in low accuracy of sleep apnea detection. At the same time, in the process of sleep apnea detection, the use of features such as breathing depth and breathing area ratio to distinguish different sleep disorder events, such as apnea, hypopnea and normal breathing, has the limitation of single feature, which makes the detection accuracy unable to be further improved. For the above reasons, we proposed an effective technical solution. Summary of the invention

[0005] 1. Technical issues to be solved

[0006] In view of the above problems, the purpose of the present invention is to provide an intelligent monitoring method for sleep apnea patients based on frequency modulated continuous wave radar. Through this radar monitoring system, the patient's sleep breathing abnormalities can be accurately monitored without contacting the human skin surface and ensuring sleep comfort. The radar sensor is used to collect data in a two-transmit and four-receive mode, and the position of the chest vibration target is located in the range-azimuth plane, so as to solve the problem of interference in the extraction of radar vital sign signals of different targets at the same distance; the respiratory signal is directly extracted by using the variational mode extraction method (WOA-VME) with whale parameter optimization, so that the main frequency narrow-band respiratory signal can be accurately extracted even when the human body's respiratory frequency state changes frequently; and the signal is processed and segmented by combining short-time Fourier transform and sliding window method to extract the time-frequency energy graph characteristics of the target radar signal, so that the network model can identify the trend of frequency changes over time and capture more comprehensive features; a hybrid neural network model is constructed by adding a bidirectional long short-term memory network module (Bi-LSTM) to a convolutional neural network (CNN) to perform four classifications of unmanned scenes, apnea, hypopnea and normal breathing conditions, so that the detection of target sleep apnea-hypopnea is more accurate; the system has the advantages of low cost, simple layout and high reliability, and can be widely used in various indoor sleeping environments.

[0007] (II) Technical solution

[0008] In order to achieve the above-mentioned purpose, the present invention specifically adopts the following technical solutions:

[0009] A sleep apnea intelligent recognition method based on frequency modulated continuous wave radar comprises the following steps:

[0010] S1. The chest vibration target positioning method of spatial two-dimensional analysis is used to extract radar sleep sign signals. The target area of ​​the extracted signal is determined by capturing the chest vibration of the human body in the sleeping state in the bed and extracting the phase information of the area, and then the sleep sign signal is obtained through signal processing;

[0011] S2. Whale Algorithm (WOA) is added to the variational mode extraction (VME) method to adaptively optimize the empirically set parameters, so as to achieve narrow-band single mode extraction with the highest frequency component of the respiratory signal as the center frequency, and directly and accurately extract the sleep breathing signal from the sleep sign signal;

[0012] S3, using sliding window method and short-time Fourier transform (STFT) to obtain sleep breathing segment data to be input into the classification neural network;

[0013] S4. The Bi-LSTM module is introduced into the convolutional neural network (CNN) model based on the time-frequency energy spectrum of the respiratory signal as input for optimization, so that it can capture more comprehensive front-to-back time series features and improve the apnea detection accuracy of the model under different sleeping postures and different sleep stages;

[0014] Furthermore, the specific steps of S1 are as follows:

[0015] S11. Use a frequency modulated continuous wave radar with an operating frequency of 60GHz. The subject lies freely and comfortably on the test bed. The millimeter wave radar sensor is placed on a high wall or ceiling to better detect the body's vital signs around the bed. The frequency modulated continuous wave radar sensor emits millimeter wave electromagnetic waves for detection. After the radar waves are reflected by the body movement and chest cavity, the device collects the echo signal and obtains the original radar data.

[0016] S12, extract the original echo information of the millimeter wave radar, and process the original echo information in the distance angle domain. The radar frequency modulation bandwidth is 4GHz, two transmitting antennas and four receiving antennas are used, and the angle resolution is improved by time division multiplexing multiple transmission and multiple reception. The angle FFT is performed on the basis of the target distance estimation to further realize the angle estimation of multiple targets. The frequency modulated continuous wave radar transmits electromagnetic waves on two transmitting (TX) antennas in a time division multiplexing (TDM) manner; and receives through four receiving (RX) channels, effectively forming 8 pairs of virtual arrays, each virtual channel corresponding to a combination of equivalent antenna receiving signals. Thus, the MIMO configuration is applied to achieve better azimuth resolution; the distance FFT is performed on each of these transmitting and receiving link (TX-RX) pairs to obtain the radar echo distance dimension data, which can separate objects occupying different range gates from each other. For each range gate of the distance dimension data, a second FFT is performed along the antenna pair (TX-RX) dimension to obtain the distance-azimuth plane, so that objects that fall within the same distance range but at different angles can be separated;

[0017] S13. Obtain a signal matrix S(i,j,t) in the range-azimuth plane from the radar system, where i represents a range unit, j represents an azimuth unit, and t represents time. For each signal at a range i and azimuth j position, extract the phase φ(i,j,t) at multiple time points;

[0018] S14. Calculate the phase mean in the time dimension The vector mean cancellation calculation of the original signal is performed using the following formula (1):

[0019]

[0020] After phase cancellation, the signal matrix S ′The dynamically changing target (chest vibration signal caused by breathing and heartbeat) in (i, j, t) is retained, and the stable background noise is suppressed;

[0021] S15, determining the region of the vibrating target using a unit average constant false alarm detection (CA-CFAR) algorithm on a range-azimuth plane to filter out reflections of static objects and further extract phase information of the target region;

[0022] S16, demodulate the I / Q orthogonal two-channel complex data collected by the target position radar using an extended DACM algorithm, expand the DACM algorithm, and add an integral process to suppress the high-frequency noise caused by the differentiator in the DACM algorithm. The extended DACM algorithm uses the following discrete form expression (2):

[0023]

[0024] The extended DACM demodulation method is used to recover the chest vibration phase of the detected target area and obtain the human sleep vital sign signal;

[0025] Furthermore, the specific steps of S2 are as follows:

[0026] S21. Directly separate the desired modal breathing signal u(t) and the residual signal s(t) from the sleep sign signal, so that the breathing signal converges at the center frequency and the residual signal energy is minimized at the frequency band where the breathing signal is located. The desired breathing signal u(t) is extracted by minimizing the following standard formula:

[0027]

[0028] u(t)+s(t)=S(t) (4) Where α is the balance factor; t is time; δ(t) is the Dirac distribution function; j is a complex unit; γ(t) is the impulse response function of the filter;

[0029] S22, establish the following augmented Lagrangian function L(u,ω d ,s,λ) to solve the variational mode extraction minimization formula (3)

[0030]

[0031] Where λ is the Lagrange multiplier; ω is the frequency; and is the frequency domain signal after Fourier transform;

[0032] S23, use the alternating direction method of multipliers (ADMM) to solve the parameters of formula (5), and iteratively update the expected breathing signal and its center frequency through ADMM formula (6) and formula (7)

[0033]

[0034] Where k is the current iteration number. Finally, the update equation (8) of the Lagrange multiplier is obtained by the dual ascent method:

[0035]

[0036] Where τ is the update parameter;

[0037] S24, the initial center frequency w of the input parameter of variational mode extraction d The balance parameter α is used as a position function and substituted into the population-based whale optimization algorithm to obtain the optimal parameters. The Shannon entropy value of the expected breathing signal is constructed as the fitness function. This fitness function is used as the iterative indicator of the whale optimization algorithm to continuously minimize and realize adaptive iterative optimization. The optimal positions of the leading whale and the remaining individuals are updated through the three predation methods of the whale algorithm: spiral encirclement predation, bubble net predation, and search predation after capture. As the iteration proceeds, the optimal initial center frequency w is continuously approached. d and equilibrium parameter α;

[0038] S25. The WOA-VME-based breathing extraction method designed above is used to directly extract a single narrow-band breathing signal component from the complex radar sleep sign signal, thereby realizing accurate extraction of sleep breathing signals.

[0039] Furthermore, the specific steps of S3 are as follows:

[0040] S31, using a fixed sliding window of 15 seconds (375 points) to intercept the sleep breathing signal, with each sliding step of 1 second (25 points), to obtain one-dimensional feature data of the breathing signal;

[0041] S32, set the sampling frequency to 25 Hz, select a spectrum window length of 64 sampling points, and the overlap length is the spectrum window length minus 1, that is, each calculation window has 63 data points overlapped. Apply Kaiser window (Kaiser) for windowing processing, and the shape factor of Kaiser window β is set to 3 to balance the resolution of time domain and frequency domain;

[0042] S33, perform fast Fourier transform (FFT), and set the FFT length to 100 times the window length to improve the frequency resolution. The specific calculation of short-time Fourier transform (STFT) can be seen in formula (9) as:

[0043]

[0044] Where x(t) is the original signal, w(τ) is the window function, τ and ω are the time and frequency indices respectively;

[0045] S34, further processing the one-dimensional respiratory characteristic data obtained in the sleep environment by a time-frequency analysis method based on short-time Fourier transform to obtain two-dimensional respiratory time-frequency energy spectrum data, which contains time-varying frequency characteristics and is more conducive to the identification and processing of sleep apnea;

[0046] Furthermore, the specific steps of S4 are as follows:

[0047] S41. Based on the data processed by S34, a CNN network model based on time-frequency spectrum input is designed. The network processes the input data through three feature extraction modules. Each feature extraction module consists of a convolution layer, an activation layer (ReLU), and a pooling layer. The convolution layer is responsible for capturing local features, the activation layer introduces nonlinear functions to learn complex features, and the pooling layer is responsible for dimensionality reduction and feature compression. In this way, the feature hierarchy is extracted and deepened;

[0048] S42, CNN network model design input sample size is 128*128*3, the three convolution layers have 32, 64 and 128 convolution kernels respectively, the convolution kernel size is also 3*3, the moving step is 1, and the training method adopts the stochastic gradient descent method which is more commonly used in deep learning;

[0049] S43. Add a bidirectional long short-term memory module after the three convolutional layers to optimize the CNN, thus forming a CNN+Bi-LSTM model. This model sends the flattened one-dimensional features of the CNN output to the Bi-LSTM module for secondary processing to enhance the model's capture of forward and backward time series features. Use two fully connected layers with ReLU activation function to further integrate the high-dimensional features of the Bi-LSTM output. Finally, the length of the Softmax output layer is 4, corresponding to 4 categories, which respectively identify and classify the four situations of unmanned scenes, apnea, hypopnea, and normal breathing;

[0050] S44. In order to ensure better learning and training effects, the cross entropy loss is combined with regularization to form a hybrid loss function that includes data loss and model complexity, which improves the generalization ability of the model on unseen data and reduces the risk of overfitting. The calculation expression of the cross entropy loss function is shown in formula (10):

[0051]

[0052] Where C is the number of categories, y i,c is a one-hot encoding vector, indicating whether sample i belongs to category C. On the basis of the cross entropy loss function, the sum of the absolute values ​​of the model weights and the sum of the squares of the model weights are added as penalty terms to reduce the model complexity and prevent overfitting. The expression of the mixed loss function is as follows:

[0053]

[0054] where λ is a hyperparameter of the regularization strength, ω j is the weight of the model. 1 and λ 2 is the regularization strength parameter, which is used to balance the weight of data loss and regularization term in the loss function. By calculating the mixed loss function, the gap between the predicted result and the true label is measured, the model parameter update is guided, and the model performance is optimized;

[0055] (III) Beneficial effects

[0056] Compared with the prior art, the present invention provides a sleep apnea intelligent recognition method based on a hybrid neural network, which has the following beneficial effects:

[0057] The present invention realizes a sleep apnea detection method based on a hybrid neural network. As a classification and recognition method using a deep learning network model, it can more accurately identify and detect sleep breathing conditions under different sleeping positions; it uses a frequency-modulated continuous wave radar sensor for detection, which simplifies and lightweights the hardware while having the ability to accurately obtain information under conditions of weak light or even no light, or when obstacles such as bedding block the equipment; the present invention realizes adaptive narrow-band respiratory signal extraction under different human respiratory differences and different respiratory frequency fluctuations, and obtains the sleep respiratory signal of the tested person as realistically as possible.

[0058] The system invention has high detection accuracy and can realize real-time, long-term, non-contact monitoring. It can be used on a large scale in places such as home bedrooms, hospitals and nursing homes, making it easy for doctors to accurately grasp the patient's condition. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in the field, other drawings can be obtained based on these drawings without creative labor.

[0060] Figure 1 It is a flow chart of the system method of the present invention;

[0061] Figure 2 A schematic flow chart of a sleep apnea intelligent recognition method based on radar respiratory signal extraction and CNN+Bi-LSTM neural network provided by the present invention;

[0062] Figure 3 A schematic diagram of the FMCW radar modulation system provided by the present invention;

[0063] Figure 4A schematic diagram of the FMCW radar transmission signal and the echo signal provided by the present invention;

[0064] Figure 5 The flowchart of the algorithm for optimizing VME parameters using the whale algorithm used in the present invention; DETAILED DESCRIPTION

[0065] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0066] The present invention will be further described below in conjunction with the embodiments and the accompanying drawings.

[0067] Example

[0068] like Figure 1 and Figure 2 The present invention provides a sleep apnea intelligent recognition method based on a hybrid neural network, and its implementation method is as follows:

[0069] S1. The chest vibration target positioning method of spatial two-dimensional analysis is used to extract radar sleep sign signals. The target area of ​​the extracted signal is determined by capturing the chest vibration of the human body in the sleeping state in the bed and extracting the phase information of the area, and then the sleep sign signal is obtained through signal processing;

[0070] S11. Use a frequency modulated continuous wave radar with an operating frequency of 60GHz. The subject lies freely and comfortably on the test bed. The millimeter wave radar sensor is placed on a high wall or ceiling to better detect the body's vital signs around the bed. The frequency modulated continuous wave radar sensor emits millimeter wave electromagnetic waves for detection. After the radar waves are reflected by the body movement and chest cavity, the device collects the echo signal and obtains the original radar data.

[0071] S12, extract the original echo information of the millimeter wave radar, and process the original echo information in the distance angle domain. The radar frequency modulation bandwidth is 4GHz, two transmitting antennas and four receiving antennas are used, and the angle resolution is improved by time division multiplexing multiple transmission and multiple reception. The angle FFT is performed on the basis of the target distance estimation to further realize the angle estimation of multiple targets. The frequency modulated continuous wave radar transmits electromagnetic waves on two transmitting (TX) antennas in a time division multiplexing (TDM) manner; and receives through four receiving (RX) channels, effectively forming 8 pairs of virtual arrays, each virtual channel corresponding to a combination of equivalent antenna receiving signals. Thus, the MIMO configuration is applied to achieve better azimuth resolution; the distance FFT is performed on each of these transmitting and receiving link (TX-RX) pairs to obtain the radar echo distance dimension data, which can separate objects occupying different range gates from each other. For each range gate of the distance dimension data, a second FFT is performed along the antenna pair (TX-RX) dimension to obtain the distance-azimuth plane, so that objects that fall within the same distance range but at different angles can be separated;

[0072] S13. Obtain a signal matrix S(i,j,t) in the range-azimuth plane from the radar system, where i represents a range unit, j represents an azimuth unit, and t represents time. For each signal at a range i and azimuth j position, extract the phase φ(i,j,t) at multiple time points;

[0073] S14. Calculate the phase mean in the time dimension The vector mean cancellation calculation of the original signal is performed using the following formula (1):

[0074]

[0075] After phase cancellation, the signal matrix S ′ The dynamically changing target (chest vibration signal caused by breathing and heartbeat) in (i, j, t) is retained, and the stable background noise is suppressed;

[0076] S15, determining the region of the vibrating target using a unit average constant false alarm detection (CA-CFAR) algorithm on a range-azimuth plane to filter out reflections of static objects and further extract phase information of the target region;

[0077] S16, demodulate the I / Q orthogonal two-channel complex data collected by the target position radar using an extended DACM algorithm, expand the DACM algorithm, and add an integral process to suppress the high-frequency noise caused by the differentiator in the DACM algorithm. The extended DACM algorithm uses the following discrete form expression (2):

[0078]

[0079] The extended DACM demodulation method is used to recover the chest vibration phase of the detected target area and obtain the human sleep vital sign signal;

[0080] S2. Whale Algorithm (WOA) is added to the variational mode extraction (VME) method to adaptively optimize the empirically set parameters, so as to achieve narrow-band single mode extraction with the highest frequency component of the respiratory signal as the center frequency, and directly and accurately extract the sleep breathing signal from the sleep sign signal;

[0081] S21. Directly separate the desired modal breathing signal u(t) and the residual signal s(t) from the sleep sign signal, so that the breathing signal converges at the center frequency and the residual signal energy is minimized at the frequency band where the breathing signal is located. The desired breathing signal u(t) is extracted by minimizing the following standard formula:

[0082]

[0083] u(t)+s(t)=S(t) (4)where α is the balance factor; t is time; δ(t) is the Dirac distribution function; j is a complex unit; γ(t) is the impulse response function of the filter;

[0084] S22, establish the following augmented Lagrangian function L(u,ω d ,s,λ) to solve the variational mode extraction minimization formula (3)

[0085]

[0086] Where λ is the Lagrange multiplier; ω is the frequency; and is the frequency domain signal after Fourier transform;

[0087] S23, use the alternating direction method of multipliers (ADMM) to solve the parameters of formula (5), and iteratively update the expected breathing signal and its center frequency through ADMM formula (6) and formula (7)

[0088]

[0089] Where k is the current iteration number. Finally, the update equation (8) of the Lagrange multiplier is obtained by the dual ascent method:

[0090]

[0091] Where τ is the update parameter;

[0092] S24, the initial center frequency w of the input parameter of variational mode extraction dThe balance parameter α is used as a position function and substituted into the population-based whale optimization algorithm to obtain the optimal parameters. The Shannon entropy value of the expected breathing signal is constructed as the fitness function. This fitness function is used as the iterative indicator of the whale optimization algorithm to continuously minimize and realize adaptive iterative optimization. The optimal positions of the leading whale and the remaining individuals are updated through the three predation methods of the whale algorithm: spiral encirclement predation, bubble net predation, and search predation after capture. As the iteration proceeds, the optimal initial center frequency w is continuously approached. d and equilibrium parameter α;

[0093] S25. The WOA-VME-based breathing extraction method designed above is used to directly extract a single narrow-band breathing signal component from the complex radar sleep sign signal, thereby realizing accurate extraction of sleep breathing signals.

[0094] S3, using sliding window method and short-time Fourier transform (STFT) to obtain sleep breathing segment data to be input into the classification neural network;

[0095] S31, using a fixed sliding window of 15 seconds (375 points) to intercept the sleep breathing signal, with each sliding step of 1 second (25 points), to obtain one-dimensional feature data of the breathing signal;

[0096] S32, set the sampling frequency to 25 Hz, select a spectrum window length of 64 sampling points, and the overlap length is the spectrum window length minus 1, that is, each calculation window has 63 data points overlapped. Apply Kaiser window (Kaiser) for windowing processing, and the shape factor of Kaiser window β is set to 3 to balance the resolution of time domain and frequency domain;

[0097] S33, perform fast Fourier transform (FFT), and set the FFT length to 100 times the window length to improve the frequency resolution. The specific calculation of short-time Fourier transform (STFT) can be seen in formula (9) as:

[0098]

[0099] Where x(t) is the original signal, w(τ) is the window function, τ and ω are the time and frequency indices respectively;

[0100] S34, further processing the one-dimensional respiratory characteristic data obtained in the sleep environment by a time-frequency analysis method based on short-time Fourier transform to obtain two-dimensional respiratory time-frequency energy spectrum data, which contains time-varying frequency characteristics and is more conducive to the identification and processing of sleep apnea;

[0101] S4. The Bi-LSTM module is introduced into the convolutional neural network (CNN) model based on the time-frequency energy spectrum of the respiratory signal as input for optimization, so that it can capture more comprehensive front-to-back time series features and improve the apnea detection accuracy of the model under different sleeping postures and different sleep stages;

[0102] S41. Based on the data processed by S34, a CNN network model based on time-frequency spectrum input is designed. The network processes the input data through three feature extraction modules. Each feature extraction module consists of a convolution layer, an activation layer (ReLU), and a pooling layer. The convolution layer is responsible for capturing local features, the activation layer introduces nonlinear functions to learn complex features, and the pooling layer is responsible for dimensionality reduction and feature compression. In this way, the feature hierarchy is extracted and deepened;

[0103] S42, CNN network model design input sample size is 128*128*3, the three convolution layers have 32, 64 and 128 convolution kernels respectively, the convolution kernel size is also 3*3, the moving step is 1, and the training method adopts the stochastic gradient descent method which is more commonly used in deep learning;

[0104] S43. Add a bidirectional long short-term memory module after the three convolutional layers to optimize the CNN, thus forming a CNN+Bi-LSTM model. This model sends the flattened one-dimensional features of the CNN output to the Bi-LSTM module for secondary processing to enhance the model's capture of forward and backward time series features. Use two fully connected layers with ReLU activation function to further integrate the high-dimensional features of the Bi-LSTM output. Finally, the length of the Softmax output layer is 4, corresponding to 4 categories, which respectively identify and classify the four situations of unmanned scenes, apnea, hypopnea, and normal breathing;

[0105] S44. In order to ensure better learning and training effects, the cross entropy loss is combined with regularization to form a hybrid loss function that includes data loss and model complexity, which improves the generalization ability of the model on unseen data and reduces the risk of overfitting. The calculation expression of the cross entropy loss function is shown in formula (10):

[0106]

[0107] Where C is the number of categories, y i,c is a one-hot encoding vector, indicating whether sample i belongs to category C. On the basis of the cross entropy loss function, the sum of the absolute values ​​of the model weights and the sum of the squares of the model weights are added as penalty terms to reduce the model complexity and prevent overfitting. The expression of the mixed loss function is as follows:

[0108]

[0109] where λ is a hyperparameter of the regularization strength, ω j is the weight of the model. 1 and λ 2is the regularization strength parameter, which is used to balance the weight of data loss and regularization term in the loss function. By calculating the mixed loss function, the gap between the predicted result and the true label is measured, the model parameter update is guided, and the model performance is optimized;

[0110] In this embodiment, Figure 2 As shown, the millimeter-wave radar acquires data in the sleeping environment around the bed, and obtains the I / Q phase information of the target through the designed distance and angle FFT, vector mean elimination in the distance-azimuth plane, and unit average constant false alarm. The sleep sign signal is obtained by extended DACM demodulation. The sleep sign signal is extracted by an adaptive sleep breathing extraction method based on WOA-VME to obtain the sleep breathing signal. The sleep breathing signal is further segmented by sliding window, and the short-time Fourier transform is used to obtain the two-dimensional sleep breathing feature data. The radar collects unmanned sleeping scene data, normal breathing data, hypopnea data, and apnea data in different sleeping positions to construct a data set, and divides the data set into training set, validation set, and test set in a ratio of 6:2:2. The designed CNN+Bi-LSTM classification and recognition network is trained and the best network model is saved;

[0111] In this embodiment, Figure 4 As shown in the figure, using frequency modulated continuous wave radar, the subjects lie freely and comfortably on the test bed, and the millimeter wave radar sensor is placed on the ceiling facing the upper body of the subjects, collecting radar echoes from 10 volunteers in different sleeping positions. In this study, the radar signal is:

[0112]

[0113] Among them A t is the amplitude of the transmitted signal of the millimeter wave radar. In addition, f c is the signal starting frequency, B is the signal bandwidth, and T is the pulse signal. The receiving antenna obtains the reflected signals of all targets within the radar detection range. Assume that the distance between the i-th target and the radar is d i , we have:

[0114]

[0115] In this embodiment, Figure 5 As shown, the algorithm flow of the whale algorithm used to optimize the VME parameters is shown, which is specifically:

[0116] When performing VME decomposition, we first need to determine the initial center frequency w dand penalty parameter α, find the best parameters of VME through WOA, use Shannon entropy to reflect the randomness and complexity of the signal, and use Shannon entropy minimization as the fitness function to optimize the selection of VME decomposition parameters. The essence of Shannon entropy is to describe the degree of information uncertainty. The larger the value, the less information the extracted signal contains; the smaller the value, the more information the extracted signal contains. The size of Shannon entropy reflects the sparse characteristics of the modal function, that is, the VME decomposition effect.

[0117] The extracted modal signal contains more noise, the periodic breathing signal is not obvious, the signal is random and complex, and the Shannon entropy value is large. If the extracted modal signal component contains more breathing signals, the signal is random and complex, and the Shannon entropy value is small. When the whale group is in a certain position (corresponding to a set of parameter combinations w d When α and α are equal, the Shannon entropy of the modal component extracted by VME under this position condition is calculated. In order to search for the global optimal component, that is, to extract the most ideal respiratory signal from the respiratory signal pool obtained by each parameter, this design takes the minimization of entropy as the fitness value in the optimization process, and takes the minimization of the entropy of the respiratory signal as the final optimization goal.

[0118] The whale algorithm updates the optimal positions of the leading whale and other individuals through three predation methods, and continuously approaches the optimal solution as the iteration proceeds.

[0119] (1) Spiral encirclement predation. The initial position of the leader whale is used as the optimal hunting position in the current group. All whales will gradually surround the prey in a spiral around the leader whale, gradually approaching the optimal solution. The formula for position update is:

[0120] D=|CX * (j)-X(j)|(14)

[0121] X(j+1)=X * (j)-AD (15)

[0122] Where D is the distance between the individual to be solved and the random individual in the contemporary population; C is the swing factor, C = 2r, where r is a random number between 0 and 1; X * (j) is the position of the leading whale; j is the current iteration number; X(j) is the position of the remaining whales; X(j+1) is the position of the individual in the next generation group; A is the convergence factor, A=2ar-a, a=2-(2j / N d ), where N d is the maximum number of iterations.

[0123] (2) Constructing a bubble net to hunt. During the hunting process, the positions of the remaining individuals and their distance from the leading whale must be calculated. The remaining individuals will continue to spiral around the leading whale and spit out bubbles to hunt prey. The position update formula during this process is:

[0124] D′=|X * (j)-X(j)|(16)

[0125] X(j+1)=D′e hl cos (2πl)+X * (j) (17)

[0126] Where h is the logarithmic spiral constant; l is a random number in [-1, 1]; D′ is the distance between the contemporary whale individual and the leading whale.

[0127] The spiral encirclement and bubble net predation are carried out simultaneously during the predation process of whales. Therefore, the probability of both predation methods is set to 50% in this paper, and the whale position formula is updated to:

[0128]

[0129] Where p is a random number between [-1, 1].

[0130] (3) Search and prey after capture. When the whale group successfully captures food, if the convergence factor |A|>1 at this time, the individual whale will break away from the leader whale and conduct individual random search and prey. This random search mechanism can prevent the whale group algorithm from falling into the local optimum; if |A|<1, the individual whale will continue to follow the leader whale to perform spiral encirclement prey and build bubble net prey. The position update formula during the search process is:

[0131] X(j+1)=X rand -AD (19)

[0132] D=|CX rand -X(j)|(20)

[0133] Where X rand is the random hunting location of individual whales.

[0134] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A sleep apnea intelligent recognition method based on a hybrid neural network, which uses a frequency modulated continuous wave radar to sense human sleep in a surrounding bed environment, characterized in that: The following steps are involved: S1. The chest vibration target positioning method of spatial two-dimensional analysis is used to extract radar sleep sign signals. The target area of ​​the extracted signal is determined by capturing the chest vibration of the human body in the sleeping state in the bed and extracting the phase information of the area, and then the sleep sign signal is obtained through signal processing; S2. Whale Algorithm (WOA) is added to the variational mode extraction (VME) method to adaptively optimize the empirically set parameters, so as to achieve narrow-band single mode extraction with the highest frequency component of the respiratory signal as the center frequency, and directly and accurately extract the sleep breathing signal from the sleep sign signal; S3, using sliding window method and short-time Fourier transform (STFT) to obtain sleep breathing segment data to be input into the classification neural network; S4. The Bi-LSTM module is introduced into the convolutional neural network (CNN) model based on the time-frequency energy spectrum of the respiratory signal as input for optimization, so that it can capture more comprehensive front-end and back-end temporal features and improve the apnea detection accuracy of the model under different sleeping postures and different sleep stages.

2. The sleep apnea intelligent recognition method according to claim 1, characterized in that: The method for locating human chest vibration in a bed environment based on frequency modulated continuous wave radar in S1 is as follows: S11. Use a frequency modulated continuous wave radar with an operating frequency of 60GHz. The subject lies freely and comfortably on the test bed. The millimeter wave radar sensor is placed on a high wall or ceiling to better detect the body's vital signs around the bed. The frequency modulated continuous wave radar sensor emits millimeter wave electromagnetic waves for detection. After the radar waves are reflected by the body movement and chest cavity, the device collects the echo signal and obtains the original radar data. S12, extract the original echo information of the millimeter wave radar, and process the original echo information in the distance angle domain. The radar frequency modulation bandwidth is 4GHz, two transmitting antennas and four receiving antennas are used, and the angle resolution is improved by time division multiplexing multiple transmission and multiple reception. The angle FFT is performed on the basis of the target distance estimation to further realize the angle estimation of multiple targets. The frequency modulated continuous wave radar transmits electromagnetic waves on two transmitting (TX) antennas in a time division multiplexing (TDM) manner; and receives through four receiving (RX) channels, effectively forming 8 pairs of virtual arrays, each virtual channel corresponding to a combination of equivalent antenna receiving signals. Thus, the MIMO configuration is applied to achieve better azimuth resolution; the distance FFT is performed on each of these transmitting and receiving link (TX-RX) pairs to obtain the radar echo distance dimension data, which can separate objects occupying different range gates from each other. For each range gate of the distance dimension data, a second FFT is performed along the antenna pair (TX-RX) dimension to obtain the distance-azimuth plane, so that objects that fall within the same distance range but at different angles can be separated; S13. Obtain a signal matrix S(i,j,t) in the range-azimuth plane from the radar system, where i represents a range unit, j represents an azimuth unit, and t represents time. For each signal at a range i and azimuth j position, extract the phase φ(i,j,t) at multiple time points; S14. Calculate the phase mean in the time dimension The vector mean cancellation calculation of the original signal is performed using the following formula: After phase cancellation, the signal matrix S ′ The dynamically changing target (chest vibration signal caused by breathing and heartbeat) in (i, j, t) is retained, and the stable background noise is suppressed; S15, determining the region of the vibrating target using a unit average constant false alarm detection (CA-CFAR) algorithm on a range-azimuth plane to filter out reflections of static objects and further extract phase information of the target region; S16, performing extended DACM algorithm demodulation on the I / Q orthogonal two-channel complex data collected by the target position radar, and the extended DACM algorithm uses the following discrete form expression: The extended DACM demodulation method is used to recover the chest vibration phase of the detected target area and obtain the human sleep vital sign signal.

3. The sleep apnea intelligent recognition method according to claim 1, characterized in that: The improved variational modal breathing signal extraction method based on WOA in S2 is: S21. Directly separate the desired modal breathing signal u(t) and the residual signal s(t) from the sleep sign signal, so that the breathing signal converges at the center frequency and the residual signal energy is minimized at the frequency band where the breathing signal is located. The desired breathing signal u(t) is extracted by minimizing the following standard formula: u(t)+s(t)=S(t) Where α is the balance factor; t is time; δ(t) is the Dirac distribution function; j is a complex unit; γ(t) is the impulse response function of the filter; S22, establish the following augmented Lagrangian function L(u,ω d ,s,λ) to solve the variational mode extraction minimization formula Where λ is the Lagrange multiplier; ω is the frequency; and is the frequency domain signal after Fourier transform; S23, use the alternating direction method of multipliers (ADMM) to solve L(u,ω) described in S22 d ,s,λ) parameters, and iteratively update the expected breathing signal and its center frequency through the following two formulas of ADMM Where k is the current iteration number. Finally, the update equation of the Lagrange multiplier is obtained by the dual ascent method: Where τ is the update parameter; S24, the initial center frequency w of the input parameter of variational mode extraction d The balance parameter α is used as a position function and substituted into the population-based whale optimization algorithm to obtain the optimal parameters. The Shannon entropy value of the expected breathing signal is constructed as the fitness function. This fitness function is used as the iterative indicator of the whale optimization algorithm to continuously minimize and realize adaptive iterative optimization. The optimal positions of the leading whale and the remaining individuals are updated through the three predation methods of the whale algorithm: spiral encirclement predation, bubble net predation, and search predation after capture. As the iteration proceeds, the optimal initial center frequency w is continuously approached. d and equilibrium parameter α; S25. The WOA-VME-based breathing extraction method designed above is used to directly extract a single narrow-band breathing signal component from the complex radar sleep sign signal, thereby achieving accurate extraction of sleep breathing signals.

4. The sleep apnea intelligent recognition method according to claim 1, characterized in that: The method for acquiring sleep breathing characteristic data in S3 is: S31, using a fixed sliding window of 15 seconds (375 points) to intercept the sleep breathing signal, with each sliding step of 1 second (25 points), to obtain one-dimensional feature data of the breathing signal; S32, set the sampling frequency to 25 Hz, select a spectrum window length of 64 sampling points, and the overlap length is the spectrum window length minus 1, that is, each calculation window has 63 data points overlapped. Apply Kaiser window (Kaiser) for windowing processing, and the shape factor of Kaiser window β is set to 3 to balance the resolution of time domain and frequency domain; S33, perform fast Fourier transform (FFT), and set the FFT length to 100 times the window length to improve the frequency resolution. The specific calculation of short-time Fourier transform (STFT) is as follows: Where x(t) is the original signal, w(τ) is the window function, τ and ω are the time and frequency indices respectively; S34. Further process the one-dimensional respiratory characteristic data obtained in the sleep environment by a time-frequency analysis method based on short-time Fourier transform to obtain two-dimensional respiratory time-frequency energy spectrum data. The data contains time-varying frequency characteristics that are more conducive to the identification and processing of sleep apnea.

5. The sleep apnea intelligent recognition method according to claim 1, characterized in that: The sleep breathing state classification method based on hybrid neural network in S4 is: S41. Based on the data processed by S34, a CNN network model based on time-frequency spectrum input is designed. The network processes the input data through three feature extraction modules. Each feature extraction module consists of a convolution layer, an activation layer (ReLU), and a pooling layer. The convolution layer is responsible for capturing local features, the activation layer introduces nonlinear functions to learn complex features, and the pooling layer is responsible for dimensionality reduction and feature compression. In this way, the feature hierarchy is extracted and deepened; S42, CNN network model design input sample size is 128*128*3, the three convolution layers have 32, 64 and 128 convolution kernels respectively, the convolution kernel size is also 3*3, the moving step is 1, and the training method adopts the stochastic gradient descent method which is more commonly used in deep learning; S43. Add a bidirectional long short-term memory module after the three convolutional layers to optimize the CNN, thus forming a CNN+Bi-LSTM model. This model sends the flattened one-dimensional features of the CNN output to the Bi-LSTM module for secondary processing to enhance the model's capture of forward and backward time series features. Use two fully connected layers with ReLU activation function to further integrate the high-dimensional features of the Bi-LSTM output. Finally, the length of the Softmax output layer is 4, corresponding to 4 categories, which respectively identify and classify the four situations of unmanned scenes, apnea, hypopnea, and normal breathing; S44. In order to ensure better learning and training effects, the cross entropy loss is combined with regularization to form a hybrid loss function that includes data loss and model complexity, which improves the generalization ability of the model on unseen data and reduces the risk of overfitting. The cross entropy loss function is calculated as follows: Where C is the number of categories, y i,c is a one-hot encoding vector, indicating whether sample i belongs to category C. On the basis of the cross entropy loss function, the sum of the absolute values ​​of the model weights and the sum of the squares of the model weights are added as penalty terms to reduce the model complexity and prevent overfitting. The hybrid loss function expression is as follows: where λ is a hyperparameter of the regularization strength, ω j is the weight of the model. λ1 and λ2 are regularization strength parameters, which are used to balance the weight of data loss and regularization term in the loss function. By calculating the mixed loss function, the gap between the predicted result and the true label is measured, the model parameter update is guided, and the model performance is optimized.

Images