Sleep disordered breathing recognition method based on bed-based mechanics sensor

Abstract

The application discloses a sleep breathing abnormality non-inductive identification method based on a bed-based mechanical sensor, and belongs to the field of biomedical engineering, and aims to solve the problems of large interference of traditional monitoring equipment and low identification accuracy caused by the difficulty in fine separation of bed-based mechanical coupling signals. The method comprises the following steps: collecting original mechanical signals containing breathing and heartbeat depth coupling characteristics; filtering out interference and removing invalid data through multi-stage filtering and quality assessment; using a generative adversarial network to decouple the coupling signals and restore them into independent breathing and heartbeat waveforms; extracting the pathological feature vectors of the restored waveforms; and inputting the pathological feature vectors into a multi-scale feature fusion model to realize automatic classification of apnea events. The application realizes completely non-inductive non-contact monitoring, improves the separation accuracy of complex coupling signals and the robustness of pathological identification, and has high clinical application value.

Description

Technical Field

[0001] This invention belongs to the field of biomedical engineering, specifically, it relates to a non-invasive method for identifying sleep breathing abnormalities based on bed-based mechanical sensors. Background Technology

[0002] With the deep integration of the big health industry and IoT technology, sleep health monitoring has become an important component of preventive medicine and chronic disease management. Sleep apnea, a clinical condition with a high incidence and significant potential harm, requires long-term monitoring and accurate identification for assessing cardiovascular risk and improving patients' quality of life. Among these technologies, bed-based sensor-based monitoring, due to its non-contact and unconstrained advantages, has become a major research direction for achieving non-invasive identification.

[0003] In the prior art, various sleep apnea detection schemes based on non-contact or bed-based sensors have emerged. For example, patent publication number CN116401598A (hereinafter referred to as "Prior Art 1") discloses a sleep apnea detection method based on physiological signals, which is the starting point and foundation of the technical improvement in this application. It proposes to separate the cardiac impulse (BCG) signal and respiratory signal from a single human micro-vibration signal using a piezoelectric sensor and a filter bank, and extract multi-dimensional features including respiratory effort features, heart rate variability features, and cardiopulmonary coupling features, and finally output the detection results through a dendritic neural network model. However, this scheme uses a traditional frequency domain filtering separation method at the signal processing front end. Because the weak vibrations caused by breathing and heartbeat in the bed-based mechanical signal exhibit deep nonlinear coupling and dynamic superposition, traditional linear filters are difficult to separate and restore the waveforms of the two in a refined and high-fidelity manner, resulting in the separation of the other component still remaining in the separated signal. This not only limits the subsequent extraction of detailed features such as heartbeat and respiratory waveform morphology, but also affects the accuracy of the final identification results and the granularity of pathological classification.

[0004] To address the insufficient accuracy of single-signal-source detection, some existing technologies have shifted towards multi-parameter fusion approaches. For example, patent publication number CN113509169A (hereinafter referred to as "Prior Art 2") introduces an independent snoring (SN) signal, which, together with the BCG signal, constructs a two-level classification model, utilizing the characteristics of snoring to assist in confirming sleep apnea events. Similarly, patent publication number CN121512464A (hereinafter referred to as "Prior Art 3") goes a step further, employing a multimodal sensor array including a pressure sensor array and a bio-radar sensor. By extracting respiratory effort signals and respiratory airflow proxy signals separately and analyzing their synergistic relationship, it identifies specific categories of respiratory events such as obstructive and central apnea events. Patent publication number CN121176854A (hereinafter referred to as "Prior Art 4") also adopts a scheme combining bio-radar and a microphone array, and is able to distinguish between obstructive and central apnea categories.

[0005] However, while these multi-source, multi-modal solutions enhance detection capabilities by adding hardware and signal sources, they also significantly increase system complexity, cost, and data fusion processing difficulty. Furthermore, their fundamental approach bypasses rather than solves the essential problem of "how to fully extract information from a single bed-based vibration signal." In particular, none of them can directly decouple and reconstruct pure respiratory and heartbeat waveforms from the most commonly used single-bed-based piezoelectric sensor signals to support accurate identification.

[0006] In summary, existing technologies generally suffer from the following unresolved technical problems: How to accurately process the deep coupling and nonlinear superposition of breathing and heartbeat signals while maintaining system simplicity and a seamless user experience, relying solely on a single type of bed-based mechanical sensor? This would enable precise decoupling and high-fidelity reconstruction of the two physiological waveforms, and on this basis, achieve high-precision automatic identification and classification of various sleep apnea events, including obstructive, central, and mixed types, to meet the accuracy and clinical reference value requirements of long-term home health monitoring. Summary of the Invention

[0007] To address the shortcomings of the existing technology, this invention provides a non-invasive method for identifying sleep breathing abnormalities based on bed-based mechanical sensors.

[0008] The technical solution adopted in this invention is as follows:

[0009] A method for the non-observable identification of sleep apnea based on bed-based mechanical sensors includes the following steps:

[0010] S1, based on the mechanical sensor array of the bed base structure, real-time acquisition of mechanical vibration signals generated by the human body in the sleep state as the original mechanical signal;

[0011] S2, performs multi-level filtering on the acquired raw mechanical signal, and performs data quality segmentation based on the signal-to-noise ratio distribution of the signal, eliminating invalid data segments caused by large-scale limb movements;

[0012] S3 decouples the preprocessed coupled signal and restores it into independent respiratory and heartbeat waveforms, while constraining the physiological consistency of the restored waveforms.

[0013] S4. Extract time-frequency domain features from the restored respiratory waveform to identify respiratory cycle, respiratory amplitude, respiratory waveform symmetry and inspiratory-expiratory time ratio, and calculate heart rate variability features.

[0014] S5 inputs the extracted feature vectors into a multi-scale feature fusion identification model, which then automatically labels and classifies various sleep apnea events.

[0015] Furthermore, in this invention, the mechanical sensor array employs multiple high-sensitivity piezoelectric ceramic sensors, which are respectively deployed at multiple support points under the mattress. These support points correspond to the core force-bearing areas distributed across the human torso and limbs, ensuring that a predetermined number of sensors are within the effective force-bearing range under different sleeping postures.

[0016] Furthermore, in this invention, the multi-level filtering process includes using a high-pass filter with a first preset cutoff frequency to remove static load offset and low-frequency creep interference from the bed base structure, and using a low-pass filter with a second preset cutoff frequency to remove power frequency interference and environmental high-frequency noise; the data quality segmentation adopts the sliding window method, setting the window length to a first preset duration and the sliding step size to a preset step size, calculating the signal variance and energy distribution within each window, comparing the signal variance with the preset average variance of normal sleep state, and if the signal variance exceeds a preset multiple of the average variance, then the signal segment corresponding to the window is determined to be a body motion interference segment and is shielded, thereby constructing a processing sequence composed of effective physiological signal segments.

[0017] Furthermore, in this invention, in S3, the decoupling and restoration of the coupled signal is achieved through a waveform separation model; the waveform separation model includes a generator network and a discriminator network; the generator network decouples and restores the single coupled signal into independent respiratory waveforms and heartbeat waveforms, and the discriminator network constrains the physiological consistency of the restored waveforms; the generator network adopts a symmetrical architecture composed of an encoder and a decoder, the encoder extracts deep spatial feature vectors from the coupled signal through multi-layer convolution operations, and the decoder restores the feature vectors to the original signal length through multi-layer transposed convolution operations and outputs independent channels corresponding to different physiological sources.

[0018] Furthermore, in this invention, the encoder includes a preset number of one-dimensional convolutional layers, each of which is followed by a linear rectified activation function and a batch normalization layer to compress the resolution of the coupled signal layer by layer and increase the number of feature channels; the decoder includes a transposed convolutional layer corresponding to the number of encoder layers, which restores the temporal resolution through upsampling operations and finally outputs two independent signal channels; the generator network adopts a course learning strategy during the training phase, first pre-training on a synthetic dataset containing simulated respiratory and heartbeat coupled signals, and then adjusting on a real bed-based mechanical signal dataset with physiological labels after the model converges, and introducing signal distortion samples under different sleep postures into the real bed-based mechanical signal dataset to improve the model's robustness in complex postures.

[0019] Furthermore, in this invention, the discriminator network adopts a multi-scale architecture, performing authenticity judgment on the waveform output by the generator network at both the time dimension and the feature frequency dimension of the original signal; the loss function of the waveform separation model is composed of a reconstruction loss term, an adversarial loss term, and a physical constraint term. The reconstruction loss term is used to ensure the consistency of the sum of the restored waveform and the original coupled signal, the adversarial loss term is used to improve the realism of the restored waveform, and the physical constraint term is based on the physiological frequency range of breathing and heartbeat, forcing the main frequency of the generated breathing waveform to be within a first preset frequency range, and requiring the main frequency of the generated heartbeat waveform to be within a second preset frequency range; through game learning between the generator network and the discriminator network, the nonlinear superposition effect of breathing and heartbeat in the mechanical transmission process is handled.

[0020] Furthermore, in this invention, the process of extracting the respiratory waveform includes:

[0021] The analytical signal of the restored respiratory waveform is calculated using Hilbert transform, and the respiratory envelope sequence is obtained from the amplitude of the analytical signal.

[0022] An adaptive peak detection algorithm is used to identify the start and end points of each respiratory cycle in the respiratory envelope sequence, thereby determining the respiratory rate and respiratory amplitude stability.

[0023] The calculation process of the heart rate variability feature includes: extracting the peak-to-peak interval of the cardiac cycle from the restored heartbeat waveform, calculating the standard deviation of the normal cardiac cycle and the root mean square of the difference between adjacent cardiac cycles in the time domain index, and calculating the low-frequency power, high-frequency power and the ratio of the two in the frequency domain index. The heart rate variability feature is used as a reference for assessing the state of the autonomic nervous system and assisting in the determination of hypoventilation events.

[0024] Furthermore, in this invention, the identification model adopts an architecture combining convolutional neural networks and long short-term memory networks. The convolutional neural network extracts local morphological distortion features of the respiratory waveform, and the long short-term memory network captures the evolution of respiratory events over time. Combined with preset pathological judgment criteria, it achieves automatic labeling and classification of various sleep apnea events. Among these, the various sleep apnea events include obstructive sleep apnea events, central sleep apnea events, and mixed sleep apnea events.

[0025] Furthermore, in this invention, the identification model also includes multi-sensor spatial fusion logic and blood oxygen saturation decrease prediction logic. The multi-sensor spatial fusion logic determines the center of gravity position of the human body on the bed surface by analyzing the time delay and amplitude difference between different sensors in the mechanical sensor array, and dynamically adjusts the weight allocation of each sensor signal in the feature fusion process based on the center of gravity position. The blood oxygen saturation decrease prediction logic indirectly infers the potential blood oxygen fluctuation trend by analyzing the attenuation rate of high-frequency power in the heart rate variability feature during apnea, and uses the blood oxygen fluctuation trend as an auxiliary feature for pathological classification.

[0026] Furthermore, in this invention, the automatic labeling and classification method for various sleep apnea events is as follows: Continuously monitor changes in respiratory amplitude. If the current respiratory amplitude decreases by more than a first preset proportion compared to the average amplitude during a reference period, and the duration of this state exceeds a second preset duration, it is determined to be an apnea event. If the decrease in respiratory amplitude is within a preset proportion range, and is accompanied by preset fluctuations in the heart rate variability characteristics, it is determined to be a hypoventilation event. After determining that an apnea event has occurred, further analyze the chest and abdominal effort signals captured by the mechanical sensor array. If bed-base micro-vibration characteristics caused by respiratory muscle contraction are detected during the cessation of respiratory airflow, it is classified as obstructive sleep apnea. If no respiratory effort signal is detected, it is classified as central sleep apnea. If the above two characteristics alternate during the same apnea period, it is classified as mixed sleep apnea.

[0027] Compared with the prior art, the present invention has the following beneficial effects:

[0028] (1) This invention introduces a waveform separation model based on generative adversarial networks to decouple and restore the coupled signals acquired by a single bed-based mechanical sensor. This achieves high-fidelity separation of independent respiratory and heartbeat waveforms from a single-source mixed signal without the need for additional sensor hardware. Compared with the traditional frequency domain filtering separation method used in prior art 1, and the multi-sensor multimodal acquisition schemes used in prior art 2, 3, and 4, this invention fundamentally solves the long-standing technical bottleneck of deep coupling and nonlinear superposition of respiration and heartbeat in bed-based mechanical signals through a completely different technical concept—namely, through game learning between the generator and the discriminator, supplemented by a composite loss function composed of reconstruction loss, adversarial loss, and physiological frequency physical constraint terms. Prior art generally tends to avoid coupling problems by adding sensors or using traditional filtering, while this invention directly tackles this problem, enabling the acquisition of independent cardiopulmonary physiological waveforms that were previously only available from polysomnography devices using only a single bed-based sensor array.

[0029] (2) This invention achieves refined preprocessing of the original mechanical signal through the synergistic cooperation of multi-level filtering and an adaptive data quality segmentation mechanism based on signal-to-noise ratio distribution. Specifically, firstly, hardware high-pass and low-pass filters are used to remove static load offset, power frequency, and environmental noise. Then, the sliding window method is used to calculate the signal variance and energy distribution to accurately identify and eliminate motion interference segments, thus constructing a high-quality effective physiological signal segment sequence to be processed. This "hardware-software collaboration and segmented quality screening" processing strategy effectively overcomes the shortcomings of existing methods, such as low signal-to-noise ratio and numerous artifact residues under complex sleep postures and environmental interference, providing a clean and reliable input signal for the subsequent waveform decoupling model, thereby ensuring the robustness of the entire identification process in the front-end stage.

[0030] (3) This invention employs an architecture combining convolutional neural networks and long short-term memory networks, enabling the model to extract local morphological distortion features of respiratory waveforms (such as the respiratory obstruction and peak clipping pattern commonly seen before obstructive sleep apnea) through convolutional neural networks, and to capture the evolution of respiratory events in the macroscopic time dimension (such as the gradual weakening and eventual disappearance of respiratory amplitude) through long short-term memory networks. This "local-global" dual-stream feature fusion design represents a qualitative leap in temporal context awareness and fine-grained feature extraction capabilities compared to the dendritic neural network used in prior art 1, the two-level SVM cascade model used in prior art 2, and the single CNN model used in prior art 4. Combined with high-fidelity waveform restoration input, this model can achieve accurate classification of obstructive, central, and mixed sleep apnea events.

[0031] (4) This invention solves the problem of insufficient generalization ability of deep learning models in physiological signal processing by combining the course learning training strategy introduced in the waveform separation model with the multi-scale discriminator architecture. The generator network is first pre-trained on a synthetic dataset containing simulated respiratory and heartbeat coupled signals. After the model converges, it is adjusted on a real bed-based mechanical signal dataset with physiological labels, and signal distortion samples under different sleep postures are introduced into the real dataset. This training strategy effectively reduces the model's dependence on the amount of real data labeled, while giving the model strong robustness to signal distortion under complex sleep postures. At the same time, the multi-scale discriminator performs authenticity judgment on the time dimension and feature frequency dimension of the original signal, respectively, ensuring that the restored waveform is not only consistent with the real physiological signal in the time domain morphology, but also strictly conforms to the physiological constraints of respiratory (dominant frequency 0.1-0.5Hz) and heartbeat (dominant frequency 0.8-3.0Hz) in the spectral distribution. This dual design at both the training strategy and network structure levels enables the waveform restoration effect of this invention to achieve a high level in terms of fidelity, robustness, and generalization ability.

[0032] (5) This invention constructs a complete multi-dimensional pathological judgment logic system, realizing a leap from single apnea detection to refined multi-subtype classification. By continuously monitoring the changes in respiratory amplitude, the multi-variable fluctuations of heart rate variability characteristics, and the chest and abdomen micro-vibration effort signals captured by the mechanical sensor array, this invention can not only accurately determine apnea and hypoventilation events, but also realize the automatic differentiation of obstructive sleep apnea (with respiratory effort signal), central sleep apnea (without respiratory effort signal), and mixed sleep apnea (two features appear alternately).

[0033] Furthermore, by combining multi-sensor spatial fusion logic for dynamic perception and weight adjustment of the bed's center of gravity position, and predictive logic for indirectly estimating blood oxygen fluctuation trends through the high-frequency power attenuation rate of heart rate variability, this invention further enriches the auxiliary feature dimensions of pathological classification. This scheme, based on high-fidelity waveform reconstruction and integrating multi-dimensional pathological features for comprehensive judgment, achieves a better balance between system simplicity, diagnostic information richness, and clinical reference value compared to existing solutions such as prior art 2 (which can only perform normal / abnormal binary classification), prior art 4 (which can distinguish snoring categories but relies on additional audio sensors), and prior art 3 (which suffers from multimodal information redundancy). Attached Figure Description

[0034] Figure 1 This is a block diagram illustrating the principle of the identification system in an embodiment of the present invention.

[0035] Figure 2 This is a flowchart of the identification method in an embodiment of the present invention.

[0036] Figure 3 This is a flowchart of single-channel coupled signal processing in an embodiment of the present invention.

[0037] Figure 4 This is a flowchart of the automatic labeling and classification method for sleep apnea events in an embodiment of the present invention. Detailed Implementation

[0038] The present invention will be further described below with reference to the accompanying drawings and embodiments. The embodiments of the present invention include, but are not limited to, the following embodiments.

[0039] like Figure 1 As shown, this embodiment of the invention provides a non-contact method for identifying sleep breathing abnormalities based on bed-based mechanical sensors, enabling long-term monitoring of a user's sleep breathing quality in a non-contact and unconstrained manner. This invention relies on a monitoring system, and the hardware system in this embodiment mainly consists of a mechanical sensor array, a signal conditioning circuit, a central processing unit, a storage module, and a wireless communication module.

[0040] The mechanical sensor array serves as the sensing layer, employing four highly sensitive piezoelectric ceramic sensors. These sensors possess extremely high piezoelectric constants, with a charge sensitivity set at 100mV / N, enabling them to capture micron-level deformation vibrations on the bed base structure. The sensors are encapsulated within a protective shell made of polyurethane material with a high elastic modulus, and are installed at four support points under the mattress, corresponding to the core force-bearing areas of the human torso and limbs. This arrangement ensures that regardless of whether the user is lying supine, on their side, or prone, at least two sensors are within the effective force-bearing range, thereby acquiring high-quality mechanical vibration signals as the original mechanical signals.

[0041] The signal conditioning circuit is connected to the sensor array via a low-noise shielded cable and integrates a preamplifier circuit, a multi-stage filtering circuit, and a 24-bit high-precision analog-to-digital converter module. The preamplifier circuit uses a high-input-impedance instrumentation amplifier to convert the high-impedance charge signal generated by the piezoelectric sensor into a low-impedance voltage signal. The multi-stage filtering circuit includes a second-order active high-pass filter with a cutoff frequency of 0.05Hz to filter out DC offset caused by the body's static weight and low-frequency creep interference from the bed base structure; it also includes a fourth-order Butterworth low-pass filter with a cutoff frequency of 45Hz to suppress power frequency electromagnetic interference and high-frequency environmental noise. The 24-bit high-precision analog-to-digital converter module quantizes the analog voltage signal into a digital signal with a dynamic range of 144dB, ensuring the complete preservation of weak high-frequency characteristics caused by heartbeats even in large-amplitude signals caused by respiratory movements.

[0042] The central processing unit (CPU) is the core of the system's computation, employing a quad-core embedded processor with a 1.5GHz clock speed. This processor integrates a dedicated deep learning acceleration unit and is equipped with 2GB of random access memory. The processor connects to the signal conditioning circuitry via a general-purpose input / output interface (GPIO), interacts with the storage module via a serial peripheral interface, and communicates with the wireless communication module via a universal asynchronous transceiver (UAV). The wireless communication module supports the Bluetooth Low Energy 5.0 protocol for real-time transmission of identification results to the user's mobile terminal or cloud platform.

[0043] like Figure 2 As shown, the specific method by which this system achieves undetectable sleep apnea identification is as follows:

[0044] First, the system acquires raw mechanical signals. The central processing unit controls the signal conditioning circuit to synchronously sample the mechanical signals from four channels at a sampling frequency of 500Hz, acquiring the mechanical vibration signals generated by the human body during sleep in real time. The sampling frequency was selected based on rigorous Nyquist sampling law calculations. Considering that the high-frequency details of heart sound signals are typically distributed between 20Hz and 150Hz, a frequency of 500Hz provides sufficient time-domain resolution for subsequent signal decoupling without ensuring signal distortion. The acquired raw data is stored in the processor's ping-pong buffer in real time in the form of data packets.

[0045] Subsequently, the central processing unit calls the digital signal processing library to further refine the filtering of the original mechanical signal. In addition to hardware filtering, an adaptive denoising algorithm based on spectral subtraction is also applied here. By analyzing the environmental noise characteristics under no-load conditions, background noise components are subtracted from the real-time signal. The processor executes data quality assessment logic, using a sliding window method to segment the signal. The window length is set to 5 seconds, and the sliding step size is 1 second. The processor calculates the signal variance and energy distribution within each window. If the variance of a certain signal segment exceeds 5 times the average variance under normal sleep conditions, it is determined that the signal segment has been interfered with by large-amplitude movements such as turning over or getting in and out of bed, and it is marked as invalid data segment and removed to avoid artifact signals misleading pathological identification.

[0046] like Figure 3As shown, the central processing unit inputs the preprocessed single-channel coupled signal into a pre-trained generative adversarial network (GAN) model. This model consists of a generator network and a discriminator network. The generator network employs a symmetrical encoder-decoder architecture. The encoder contains five one-dimensional convolutional layers with progressively decreasing kernel size and a stride of 2, used to compress long sequences of coupled signals into high-dimensional latent space feature vectors. Each convolutional layer is followed by a batch normalization layer and a linear rectified activation function to enhance the model's nonlinear expressive power. The decoder contains five transposed convolutional layers, which upsample the latent space features to gradually restore them to the original signal length. The decoder's output layer has two independent channels, corresponding to the restored respiratory waveform and heartbeat waveform, respectively.

[0047] To ensure that the separated waveforms have genuine physiological and physical meaning, this embodiment introduces a multi-scale discrimination mechanism into the discriminator network. The discriminator not only determines the authenticity of the original signal in the time dimension but also imposes constraints in the feature frequency dimension. The training process of the generative adversarial network follows the following loss function logic:

[0048]

[0049] in, Represents the reconstruction loss, used to ensure the summation consistency between the reconstructed waveform and the original signal; its weighting coefficients... Set to 10; Representing adversarial loss, used to enhance the realism of waveforms, its weighting coefficients... Set to 1; Representing physical constraints, this requires that the dominant frequency of the respiratory waveform must be within the physiological range of 0.1Hz to 0.5Hz, and the dominant frequency of the heartbeat waveform must be within the physiological range of 0.8Hz to 3.0Hz. This decoupling method based on generative adversarial networks can effectively handle the nonlinear superposition problem of respiration and heartbeat in the mechanical transmission process.

[0050] Next, the system extracts pathological features and calculates parameters. For the reconstructed high-fidelity respiratory waveform, the central processing unit uses Hilbert transform to calculate its analytic signal, thereby extracting the respiratory envelope sequence. Through an adaptive peak detection algorithm, the processor can accurately locate the inspiratory start, expiratory start, and respiratory end of each respiratory cycle, and then calculate the respiratory rate, respiratory amplitude stability, and inspiratory-expiratory time ratio. For the reconstructed heartbeat waveform, the processor extracts the peak-to-peak interval of the cardiac cycle and calculates heart rate variability characteristics, including the standard deviation of a normal cardiac cycle, the root mean square of the difference between adjacent cardiac cycles, and the ratio of low-frequency power to high-frequency power in the frequency domain. These characteristics will serve as important references for assessing the state of the autonomic nervous system.

[0051] Subsequently, the central processing unit inputs the extracted feature vectors into a fusion recognition model composed of a convolutional neural network and a long short-term memory network. The convolutional neural network is responsible for extracting the morphological distortion features of the respiratory waveform at the microscale, such as the respiratory obstruction and peak clipping features commonly seen before obstructive sleep apnea occurs; the long short-term memory network is responsible for capturing the evolution of the signal in the macroscopic time dimension, such as the process of the respiratory amplitude gradually decreasing until it disappears.

[0052] In this embodiment, as Figure 4 As shown, the automatic labeling and classification methods for various sleep apnea events are as follows: The processor continuously monitors changes in respiratory amplitude. If the current respiratory amplitude decreases by more than 90% compared to the average amplitude over the past 120 seconds, and this state lasts for 10 seconds or more, it is determined to be an apnea event. If the amplitude decrease is between 30% and 90%, and is accompanied by significant fluctuations in heart rate variability, it is determined to be a hypoventilation event. After confirming the occurrence of apnea, the processor further analyzes the weak chest and abdominal effort signals captured by the mechanical sensors. If micro-vibrations of the bed base caused by respiratory muscle contraction can still be detected during the cessation of airflow, it is classified as obstructive sleep apnea. If no respiratory effort signal is detected, it is classified as central sleep apnea. If the characteristics of both types alternate, it is classified as mixed type.

[0053] After the monitoring task is completed, the central processing unit summarizes the data from the entire day and calculates the apnea-hypopnea index. This refers to the total number of apnea and hypoventilation events occurring per hour. The calculation formula is as follows:

[0054]

[0055] in, The total number of apnea events. This represents the total number of low ventilation events. This represents the total effective sleep hours. Based on the calculation results, the system classifies sleep quality into four levels: normal (index less than 5), mild (5 to 15), moderate (15 to 30), and severe (greater than 30), and sends detailed sleep trend charts to the user's mobile phone via Bluetooth module.

[0056] In summary, this invention achieves completely undetectable and highly accurate sleep apnea identification by integrating a high-sensitivity mechanical sensor array into the bed base structure, combined with a waveform decoupling algorithm based on generative adversarial networks and a deep learning identification model that fuses multi-scale features. It not only solves the problem of inconvenience in wearing traditional devices, but also overcomes the technical bottlenecks of signal coupling and environmental interference in non-contact monitoring through advanced signal processing methods, providing a reliable technical solution for the early screening and home management of chronic respiratory diseases.

[0057] The above embodiments are merely one of the preferred embodiments of the present invention and should not be used to limit the scope of protection of the present invention. Any modifications or refinements made to the main design concept and spirit of the present invention that are not of substantial significance, but solve the same technical problem as the present invention, should be included within the scope of protection of the present invention.

Claims

1. A method for non-invasive identification of sleep breathing abnormalities based on bed-based mechanical sensors, characterized in that, Includes the following steps: S1, based on the mechanical sensor array of the bed base structure, real-time acquisition of mechanical vibration signals generated by the human body in the sleep state as the original mechanical signal; S2, performs multi-level filtering on the acquired raw mechanical signal, and performs data quality segmentation based on the signal-to-noise ratio distribution of the signal, eliminating invalid data segments caused by large-scale limb movements; S3, the preprocessed coupled signal is decoupled and restored to independent respiratory and heartbeat waveforms, while constraining the physiological consistency of the restored waveforms; wherein, the decoupling and restoration of the coupled signal is achieved through a waveform separation model; the waveform separation model includes a generator network and a discriminator network; the generator network decouples the single coupled signal and restores it to independent respiratory and heartbeat waveforms, and the discriminator network constrains the physiological consistency of the restored waveforms; the generator network adopts a symmetrical architecture composed of an encoder and a decoder, the encoder extracts deep spatial feature vectors from the coupled signal through multi-layer convolution operations, and the decoder restores the feature vectors to the original signal length through multi-layer transposed convolution operations and outputs independent channels corresponding to different physiological sources; S4. Extract time-frequency domain features from the restored respiratory waveform to identify respiratory cycle, respiratory amplitude, respiratory waveform symmetry and inspiratory-expiratory time ratio, and calculate heart rate variability features. S5 inputs the extracted feature vectors into a multi-scale feature fusion identification model, which then automatically labels and classifies various sleep apnea events.

2. The method for non-invasive identification of sleep breathing abnormalities based on bed-based mechanical sensors according to claim 1, characterized in that, The mechanical sensor array employs multiple high-sensitivity piezoelectric ceramic sensors, which are deployed at multiple support points under the mattress. These support points correspond to the core force-bearing areas of the human torso and limbs, ensuring that a predetermined number of sensors are within the effective force-bearing range under different sleeping postures.

3. The method for non-invasive identification of sleep breathing abnormalities based on bed-based mechanical sensors according to claim 1, characterized in that, The multi-level filtering process includes using a high-pass filter with a first preset cutoff frequency to remove static load offset and low-frequency creep interference from the bed base structure, and using a low-pass filter with a second preset cutoff frequency to remove power frequency interference and high-frequency environmental noise. The data quality segmentation adopts the sliding window method, setting the window length to a first preset duration and the sliding step size to a preset step size. The signal variance and energy distribution within each window are calculated, and the signal variance is compared with the preset average variance of normal sleep state. If the signal variance exceeds a preset multiple of the average variance, the signal segment corresponding to the window is determined to be a body motion interference segment and is shielded, thereby constructing a processing sequence composed of effective physiological signal segments.

4. The method for non-invasive identification of sleep breathing abnormalities based on bed-based mechanical sensors according to claim 1, characterized in that, In S3, the encoder includes a preset number of one-dimensional convolutional layers. Each one-dimensional convolutional layer is followed by a linear rectified activation function and a batch normalization layer, which are used to compress the resolution of the coupled signal layer by layer and increase the number of feature channels. The decoder contains transposed convolutional layers corresponding to the number of encoder layers. It restores temporal resolution through upsampling operations and finally outputs two independent signal channels. The generator network adopts a course learning strategy during the training phase. It is first pre-trained on a synthetic dataset containing simulated respiratory and heartbeat coupling signals. After the model converges, it is adjusted on a real bed-based mechanical signal dataset with physiological labels. Signal distortion samples under different sleep postures are introduced into the real bed-based mechanical signal dataset to improve the model's robustness in complex postures.

5. The method for non-invasive identification of sleep breathing abnormalities based on bed-based mechanical sensors according to claim 4, characterized in that, The discriminator network adopts a multi-scale architecture, which performs authenticity judgment on the waveform output by the generator network on both the time dimension and the feature frequency dimension of the original signal. The loss function of the waveform separation model is composed of a reconstruction loss term, an adversarial loss term, and a physical constraint term. The reconstruction loss term is used to ensure the consistency of the sum of the restored waveform and the original coupled signal. The adversarial loss term is used to improve the realism of the restored waveform. The physical constraint term is based on the physiological frequency range of breathing and heartbeat, and forces the main frequency of the generated breathing waveform to be within a first preset frequency range and the main frequency of the generated heartbeat waveform to be within a second preset frequency range. The nonlinear superposition effect of respiration and heartbeat in the mechanical transmission process is handled through game-like learning between the generator network and the discriminator network.

6. The method for non-invasive identification of sleep breathing abnormalities based on bed-based mechanical sensors according to claim 1, characterized in that, The process of extracting the respiratory waveform includes: The analytical signal of the restored respiratory waveform is calculated using Hilbert transform, and the respiratory envelope sequence is obtained from the amplitude of the analytical signal. An adaptive peak detection algorithm is used to identify the start and end points of each respiratory cycle in the respiratory envelope sequence, thereby determining the respiratory rate and respiratory amplitude stability. The calculation process of the heart rate variability feature includes: extracting the peak-to-peak interval of the cardiac cycle from the restored heartbeat waveform, calculating the standard deviation of the normal cardiac cycle and the root mean square of the difference between adjacent cardiac cycles in the time domain index, and calculating the low-frequency power, high-frequency power and the ratio of the two in the frequency domain index. The heart rate variability feature is used as a reference for assessing the state of the autonomic nervous system and assisting in the determination of hypoventilation events.

7. The method for non-invasive identification of sleep breathing abnormalities based on bed-based mechanical sensors according to claim 1, characterized in that, The identification model adopts an architecture combining convolutional neural networks and long short-term memory networks. The convolutional neural network extracts local morphological distortion features of respiratory waveforms, while the long short-term memory network captures the evolution of respiratory events over time. Combined with preset pathological judgment criteria, it achieves automatic labeling and classification of various sleep apnea events. Among these, various sleep apnea events include obstructive sleep apnea events, central sleep apnea events, and mixed sleep apnea events.

8. The method for non-invasive identification of sleep breathing abnormalities based on bed-based mechanical sensors according to claim 1, characterized in that, The identification model also includes multi-sensor spatial fusion logic and blood oxygen saturation decline prediction logic. The multi-sensor spatial fusion logic determines the center of gravity position of the human body on the bed by analyzing the time delay and amplitude difference between different sensors in the mechanical sensor array, and dynamically adjusts the weight allocation of each sensor signal in the feature fusion process based on the center of gravity position. The blood oxygen saturation decline prediction logic indirectly infers the potential blood oxygen fluctuation trend by analyzing the decay rate of high-frequency power in the heart rate variability feature during apnea, and uses the blood oxygen fluctuation trend as an auxiliary feature for pathological classification.

9. The method for non-invasive identification of sleep breathing abnormalities based on bed-based mechanical sensors according to claim 1, characterized in that, The automatic labeling and classification method for various sleep apnea events is as follows: continuously monitor the changes in respiratory amplitude. If the current respiratory amplitude decreases by more than a first preset proportion compared to the average amplitude of the reference period, and the duration of this state exceeds a second preset duration, it is determined to be an apnea event. If the decrease in respiratory amplitude is within a preset proportion range, and is accompanied by preset fluctuations in the heart rate variability characteristics, it is determined to be a hypoventilation event. After determining that a breathing apnea event has occurred, the chest and abdominal effort signals captured by the mechanical sensor array are further analyzed. If bed-base micro-vibration characteristics caused by respiratory muscle contraction are detected during the cessation of airflow, it is classified as obstructive sleep apnea; if no breathing effort signal is detected, it is classified as central sleep apnea; if the above two characteristics alternate during the same breathing apnea period, it is classified as mixed sleep apnea.

Images