CN-121971036-A - Sleep stage method and system based on two-way non-contact sensing and deep learning

CN121971036ACN 121971036 ACN121971036 ACN 121971036ACN-121971036-A

Abstract

The invention relates to the field of physiological signal processing and deep learning, and discloses a sleep stage method and a sleep stage system based on double-path non-contact sensing and deep learning, wherein a non-contact type sensor is arranged below a mattress to collect multi-source signals by utilizing a non-contact type multi-channel body movement signal collecting module, and the multi-source signals are fused; the method comprises the steps of utilizing a multi-mode signal preprocessing module to call multi-source signals of a non-contact multi-channel body movement signal acquisition module, carrying out synchronous, denoising, artifact correction and segmentation standardization processing on the multi-source signals, utilizing a deep learning sleep stage module to call signals processed by the multi-mode signal preprocessing module, adopting a deep learning network to extract and classify space-time characteristics of the processed signals, establishing an end-to-end deep learning model, utilizing a sleep stage output and analysis module to call the deep learning model established by the deep learning sleep stage module, generating a sleep structure diagram and sleep quality index through the deep learning model, and carrying out abnormal early warning.

Inventors

ZHANG SHILI
GU YUWEI
WU YANHUI

Assignees

宁波数字孪生(东方理工)研究院

Dates

Publication Date: 20260505
Application Date: 20251218

Claims (10)

1. The sleep stage method based on the double-path non-contact sensing and the deep learning is characterized by being applied to a sleep stage system based on the double-path non-contact sensing and the deep learning, wherein the system comprises a non-contact multi-channel body movement signal acquisition module, a multi-mode signal preprocessing module, a deep learning sleep stage module and a sleep stage output and analysis module which are in communication connection with each other, and the method comprises the following steps: Step S1, a non-contact sensor is arranged below a mattress by utilizing the non-contact multi-channel body movement signal acquisition module to acquire multi-source signals, and the multi-source signals are fused; s2, calling the multi-source signals of the non-contact multi-channel body motion signal acquisition module by utilizing the multi-mode signal preprocessing module, and carrying out synchronization, denoising, artifact correction and segmentation standardization processing on the multi-source signals; S3, calling the signals processed by the multi-mode signal preprocessing module by using the deep learning sleep stage module, extracting and classifying space-time characteristics of the processed signals by using a deep learning network, and establishing an end-to-end deep learning model; And S4, calling a deep learning model established by the deep learning sleep stage output and analysis module, generating a sleep structure diagram and a sleep quality index through the deep learning model, and carrying out abnormality early warning.
2. The sleep staging method based on two-way non-contact sensing and deep learning according to claim 1, characterized in that the non-contact multichannel body movement signal acquisition module includes a piezoelectric body movement signal acquisition unit including a piezoelectric film sensor disposed under a mattress, the step S1 includes: And S11, acquiring pressure fluctuation of the back of the human body and fine mechanical vibration caused by heart pulsation and respiratory motion by using the piezoelectric body motion signal acquisition unit through the piezoelectric film sensor.
3. The sleep staging method based on two-way non-contact sensing and deep learning according to claim 2, characterized in that the non-contact multichannel body movement signal acquisition module includes a piezoresistive body movement signal acquisition unit including a piezoresistive effect sensor array arranged under a mattress, the step S1 includes: and step S12, detecting static contact pressure distribution between the body and the bed surface by using the piezoresistance body movement signal acquisition unit through the piezoresistance effect sensor, and monitoring body position transformation, respiratory rhythm and sleeping behavior characteristics.
4. The sleep staging method based on two-way non-contact sensing and deep learning according to claim 3, wherein the non-contact multi-channel body movement signal acquisition module includes a multi-source signal fusion and anti-interference processing unit, the step S1 includes: And S13, synchronously acquiring signals of the piezoelectric body motion signal acquisition unit and the piezoresistance body motion signal acquisition unit by utilizing a signal conditioning circuit and an edge end preprocessing algorithm integrated in the multi-source signal fusion and anti-interference processing unit, performing noise suppression, drift compensation and characteristic enhancement, realizing stable extraction of respiratory waveform, heart mechanical vibration and body motion event through fusion time sequence alignment, and enhancing robustness under different body types, sleeping postures and mattress conditions.
5. The sleep staging method based on two-way non-contact sensing and deep learning according to claim 1, characterized in that step S2 includes: Step S21, aligning the received multi-source signals in time sequence based on hardware triggering and a time stamping mechanism; S22, restraining power frequency interference, environmental noise and baseline drift by adopting a band-pass filtering mode, an empirical mode decomposition mode and a wavelet noise reduction mode, and marking and correcting a severe motion pseudo film section by combining a body motion intensity detection result; Step S23, segmenting and normalizing the preprocessed signals by taking 30 seconds as a time slice, and providing an input sequence with a unified format for the deep learning model.
6. The sleep staging method based on two-way non-contact sensing and deep learning according to claim 1, wherein step S3 includes: s31, establishing the deep learning model, and respectively extracting the structural characteristics of local waveforms of all modes through a one-dimensional convolutional neural network by taking signal streams of different modes as input; Step S32, modeling a long-range dependency relationship of a time sequence by utilizing a bidirectional gating circulating unit network; Step S33, introducing a multi-head self-attention mechanism in a feature fusion stage, and adaptively learning the relative importance of each mode signal in different sleep stages; And step S34, the output end of the deep learning model adopts a Softmax classification layer, and probability distribution of wakefulness, shallow sleep, deep sleep and rapid eye movement period is given to each 30-second time slice.
7. The sleep staging method based on two-way non-contact sensing and deep learning according to claim 1, characterized in that the sleep staging output and analysis module is in communication connection with a user terminal capable of displaying sleep detection and analysis results, the step S4 includes: S41, the sleep stage output and analysis module encrypts and uploads sleep stage results and statistical data generated by reasoning to a cloud server through a local gateway or an integrated Wi-Fi module for persistent storage and long-range trend analysis; Step S42, the sleep stage output and analysis module is used for continuously monitoring the apnea event in the sleep process by combining with the breathing waveform characteristics, evaluating the apnea low-ventilation index risk level, analyzing whether the sleep period is complete, whether the deep sleep is absent or fragmented, identifying the sleep structure disorder condition, and pushing personalized sleep improvement advice, environment adjustment scheme or medical reminding to the user.
8. The sleep stage system based on the two-way non-contact sensing and the deep learning is characterized by comprising a non-contact multi-channel body movement signal acquisition module, a multi-mode signal preprocessing module, a deep learning sleep stage module and a sleep stage output and analysis module which are connected with each other in a communication way, The non-contact type multichannel body movement signal acquisition module is used for acquiring multi-source signals through a non-contact type sensor and carrying out fusion processing on the multi-source signals; the multi-mode signal preprocessing module is used for carrying out synchronization, denoising, artifact correction and segmentation standardization processing on the multi-source signals; the deep learning sleep stage module is used for extracting and classifying space-time characteristics of the processed signals through a deep learning network and establishing an end-to-end deep learning model; The sleep stage output and analysis module is used for generating a sleep structure diagram and sleep quality indexes through the deep learning model and carrying out abnormal early warning; The non-contact multi-channel body movement signal acquisition module synchronously acquires heart impact signals, respiratory movements and body movement behaviors of a human body through sensors arranged under a mattress, then transmits the acquired signals to the multi-mode signal preprocessing module, the multi-mode signal preprocessing module aligns the channel signals based on a hardware trigger mechanism, removes power frequency and environmental noise through a filtering algorithm, cuts the channel signals according to time windows, inputs the aligned multi-mode sequences to the deep learning sleep stage module, the deep learning sleep stage module establishes a deep learning model, extracts local waveform characteristics through one-dimensional convolution, captures long-range time sequence dependence through a bidirectional circulating network, adaptively fuses different mode characteristics through an attention mechanism, and finally outputs sleep stage probabilities corresponding to each time window according to the deep learning model by the sleep stage output and analysis module and generates a sleep report.
9. The sleep stage system based on two-way non-contact sensing and deep learning according to claim 8, wherein the non-contact multi-channel body movement signal acquisition module comprises a piezoelectric body movement signal acquisition unit, a piezoresistance body movement signal acquisition unit and a signal processing unit, The piezoelectric body motion signal acquisition unit is used for acquiring pressure fluctuation of the back of a human body and fine mechanical vibration caused by heart pulsation and respiratory motion through a piezoelectric film sensor arranged below the mattress; the piezoresistive body movement signal acquisition unit is used for detecting static contact pressure distribution between a body and a bed surface through a piezoresistive effect sensor array arranged below the mattress and monitoring body position transformation, respiratory rhythm and sleeping behavior characteristics; The signal processing unit is arranged below the mattress and is used for processing the reasoning process of the deep learning model so as to ensure the real-time performance and privacy safety of data.
10. The sleep staging system based on two-way non-contact sensing and deep learning according to claim 9, further comprising an edge device layer, a gateway layer, a cloud service layer, an application layer, The edge equipment layer comprises a sensor acquisition module and an edge calculation unit which are in communication connection with each other, wherein the sensor acquisition module is used for acquiring signals through a piezoelectric film sensor and a piezoresistive effect sensor, and carrying out ADC signal conversion and local caching; the gateway layer comprises a local intelligent gateway, wherein the local intelligent gateway is in communication connection with the edge computing unit through a Wi-Fi hot spot or a router relay, and is used for processing multi-device connection management, protocol conversion, edge thrust, message queue transmission and information encryption; The cloud service layer comprises an AI analysis platform and a data storage and management unit, wherein the AI analysis platform is used for processing deep learning reasoning, sleep stage calculation, big data analysis and GPU clusters, and the data storage and management unit is used for processing a time sequence database, object storage, user management and data encryption; The application layer comprises a mobile terminal APP and a Web management background, wherein the mobile terminal APP is suitable for an IOS or Android system and can monitor and generate a report in real time, and the Web management background is applied to a medical institution terminal and is used for multi-user management and data export.

Description

Sleep stage method and system based on two-way non-contact sensing and deep learning Technical Field The invention relates to the technical field of physiological signal processing and deep learning, in particular to a sleep stage method and a sleep stage system based on two-way non-contact sensing and deep learning. Background In the sleeping process, a series of functions of the brain, muscles, eyes, heart, breath and the like of the human body can be changed, and the judgment of the sleeping quality of the human body can be promoted by monitoring the changes, so that sleeping disorders can be diagnosed and treated in time. In the prior art, the method for detecting the sleep quality has some defects, namely 1, the characteristic engineering relies on the adoption of manual characteristics (HRV statistics, respiratory rate and body movement intensity) +a traditional classifier, the generalization capability is weak, 2, single-path signals are limited, the signal to noise ratio of single piezoelectric or piezoresistive signals in the scenes such as body position change, lateral lying, bed edge and the like is suddenly reduced, 3, shallow sleep/REM precision is poor, most non-contact schemes can only be classified in coarse granularity, the F1 value of the boundary stage N1/N2/REM is generally lower than 0.6, 4, the prior patent from end to end is mainly a sensor+manual characteristics+shallow model, and an end-to-end deep learning scheme is not formed. For example, the chinese patent with publication No. CN111150378B is based on an optical fiber sensor, but the algorithm accuracy is not improved enough, the chinese patent with publication No. CN111467644B is mainly based on body movement monitoring, and cannot reliably distinguish the complete sleep stage, the chinese patent with publication No. CN102065753a relies on attached volume pulse signals, the comfort is poor, and the chinese patent with publication No. CN107205652a does not propose a deep learning scheme for mattress signals. Therefore, the prior art lacks a complete sleep stage scheme which can fuse multi-mode mechanical physiological signals without wearing a sensor under the condition of a common mattress structure and can achieve the accuracy close to PSG in a real environment. Disclosure of Invention To solve at least one of the above problems, the present invention firstly provides a sleep stage method based on two-way non-contact sensing and deep learning, which is applied to a sleep stage system based on two-way non-contact sensing and deep learning, wherein the system comprises a non-contact multi-channel body movement signal acquisition module, a multi-mode signal preprocessing module, a deep learning sleep stage module and a sleep stage output and analysis module which are in communication connection with each other, and the method comprises: Step S1, a non-contact sensor is arranged below a mattress by utilizing the non-contact multi-channel body movement signal acquisition module to acquire multi-source signals, and the multi-source signals are fused; s2, calling the multi-source signals of the non-contact multi-channel body motion signal acquisition module by utilizing the multi-mode signal preprocessing module, and carrying out synchronization, denoising, artifact correction and segmentation standardization processing on the multi-source signals; S3, calling the signals processed by the multi-mode signal preprocessing module by using the deep learning sleep stage module, extracting and classifying space-time characteristics of the processed signals by using a deep learning network, and establishing an end-to-end deep learning model; And S4, calling a deep learning model established by the deep learning sleep stage output and analysis module, generating a sleep structure diagram and a sleep quality index through the deep learning model, and carrying out abnormality early warning. Optionally, the non-contact multichannel body motion signal acquisition module includes a piezoelectric body motion signal acquisition unit, the piezoelectric body motion signal acquisition unit includes a piezoelectric film sensor disposed under a mattress, and the step S1 includes: And S11, acquiring pressure fluctuation of the back of the human body and fine mechanical vibration caused by heart pulsation and respiratory motion by using the piezoelectric body motion signal acquisition unit through the piezoelectric film sensor. Optionally, the non-contact multichannel body motion signal acquisition module includes a piezoresistive body motion signal acquisition unit, the piezoresistive body motion signal acquisition unit includes a piezoresistive effect sensor array disposed below the mattress, and the step S1 includes: and step S12, detecting static contact pressure distribution between the body and the bed surface by using the piezoresistance body movement signal acquisition unit through the piezoresistance effect sensor, and monitoring body position tr