Sleep quality monitoring and abnormal breathing early warning system based on generative adversarial network

Abstract

The invention discloses a sleep quality monitoring and abnormal breathing early warning system based on a generative adversarial network, and particularly relates to the technical field of intelligent health monitoring. The system comprises a signal acquisition module, a data processing module, a generative adversarial network model library and an early warning module, the signal acquisition module is used for acquiring a breathing vibration signal, a body movement signal, a blood oxygen signal, a photoelectric volume pulse wave signal and a wrist body movement signal in parallel through a non-contact mattress piezoelectric sensor and an intelligent bracelet; the data processing module generates comparison data through a first processing method and a second processing method, and outputs a comprehensive feature vector after integration processing; a discriminator of the generative adversarial network model library adopts a multi-task learning architecture and outputs an abnormal index; and the early warning module realizes third-level early warning according to the abnormal index. Through multi-mode signal fusion processing and a non-contact monitoring mode, the accuracy, the reliability and the user experience of sleep breathing abnormity monitoring are effectively improved.

Description

Technical Field

[0001] This invention discloses a sleep quality monitoring and respiratory abnormality early warning system based on generative adversarial networks, specifically relating to the field of intelligent health monitoring technology. Background Technology

[0002] In the development of modern sleep health monitoring technology, the demand for non-intrusive and continuous monitoring is increasing. With rising awareness of health management, people are paying more attention to sleep quality, and traditional single-sensor-based monitoring solutions face serious challenges in terms of data integrity and accuracy. Multimodal physiological signal monitoring, with its significant advantages such as information complementarity and cross-validation, has become a key technology for achieving accurate sleep monitoring.

[0003] Traditional solutions mainly employ contact sensor monitoring or single-modal monitoring. Contact sensor monitoring collects signals such as electrocardiogram and respiration through devices such as electrode patches and chest strap respiratory monitors, requiring users to wear multiple devices, which affects sleep comfort. Single-modal monitoring relies on mattress pressure sensors or a single signal source in wearable devices, such as analyzing sleep status solely through body movement or determining respiratory events through blood oxygen saturation.

[0004] However, traditional sleep monitoring solutions have significant limitations in implementation: on the one hand, while contact monitoring devices offer high signal quality, they are uncomfortable, easily disrupting normal sleep and leading to data distortion; on the other hand, while single-modal monitoring offers some convenience, it lacks cross-validation from multiple signal sources. When faced with individual differences and complex sleep scenarios, existing systems lack effective signal quality assessment and anomaly verification mechanisms, resulting in insufficient reliability and practicality of monitoring results. Furthermore, in terms of respiratory anomaly identification, traditional threshold judgment methods struggle to balance sensitivity and specificity, leading to a high false alarm rate and inability to achieve accurate early warning. Therefore, developing a sleep quality monitoring and respiratory anomaly early warning system that enables non-intrusive monitoring, multi-modal signal fusion capabilities, and intelligent anomaly identification is of great significance. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a sleep quality monitoring and breathing abnormality early warning system based on generative adversarial networks to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a sleep quality monitoring and respiratory abnormality early warning system based on generative adversarial networks, comprising:

[0007] Signal acquisition module, data processing module, generative adversarial network model library, and early warning module;

[0008] The signal acquisition module consists of a non-contact mattress piezoelectric sensor and a wearable smart bracelet, used to collect the following objective physiological signals from the user in parallel: respiratory vibration signals and body movement signals are collected through the mattress piezoelectric sensor; blood oxygen signals, photoplethysmography pulse wave signals and wrist body movement signals are collected through the smart bracelet;

[0009] The data processing module, which communicates with the signal acquisition module, is used to process the acquired multimodal objective physiological signals to generate a comprehensive feature vector.

[0010] A generative adversarial network (GAN) model library stores trained GAN models; the discriminator of the GAN model is configured to receive the comprehensive feature vector output by the data processing module and output an abnormality index representing the probability of respiratory abnormalities.

[0011] The early warning module communicates with the generative adversarial network model library and is used to compare the abnormal index with a preset threshold. When the abnormal index exceeds the preset threshold, it generates and issues a respiratory abnormality early warning signal.

[0012] The data processing module is configured to perform the following operations:

[0013] The respiratory vibration signal and the body movement signal are defined as data A;

[0014] The blood oxygen signal and the photoplethysmography signal are defined as data B;

[0015] The wrist movement signal is defined as data C;

[0016] Data A and data B are processed using the first processing method to generate comparison data 1;

[0017] Data B and data C are processed using the second processing method to generate comparison data 2;

[0018] The comparison data 1 and comparison data 2 are integrated and processed to generate data D as a comprehensive feature vector.

[0019] Preferably, the data processing module is configured to perform the following operations:

[0020] The respiratory vibration signal and the body movement signal are defined as data A;

[0021] The blood oxygen signal and the photoplethysmography signal are defined as data B;

[0022] The wrist movement signal is defined as data C;

[0023] Data A and data B are processed using the first processing method to generate comparison data 1;

[0024] Data B and data C are processed using the second processing method to generate comparison data 2;

[0025] The comparison data 1 and comparison data 2 are integrated and processed to generate data D as a comprehensive feature vector.

[0026] Preferably, the data processing module is configured to perform the following operations:

[0027] The respiratory vibration signal and body movement signal are defined as data A;

[0028] The blood oxygen signal and photoplethysmography signal are defined as data B;

[0029] The wrist movement signal is defined as data C;

[0030] Data A and data B are processed using the first processing method to generate comparison data 1;

[0031] Data B and data C are processed using the second processing method to generate comparison data 2;

[0032] The comparison data 1 and comparison data 2 are integrated and processed to generate data D as the comprehensive feature vector.

[0033] Preferably, the first processing method is to calculate the multimodal correlation between data A and data B, specifically including:

[0034] Calculate the first Pearson correlation coefficient between respiratory vibration signal and blood oxygen signal. ;

[0035] Calculate the second Pearson correlation coefficient between the body motion signal and the photoplethysmography (PPG) signal. ;

[0036] The comparison data 1 was obtained by weighting the first Pearson correlation coefficient and the second Pearson correlation coefficient.

[0037] Preferably, the second processing method is to calculate the joint time-domain variability of data B and data C, specifically obtaining comparison data 2 through the following formula:

[0038] ,

[0039] in, and These are the photoplethysmography (PPG) signal and wrist motion signal values ​​at the i-th sampling point, respectively. Let N be the mean of the product sequence of the two, and N be the number of sampling points within the time window.

[0040] Preferably, the integration process is an adaptive weighted fusion based on the signal-to-noise ratio to generate data D.

[0041] Preferably, the data processing module further includes a signal quality verification unit, used to evaluate the quality of the original signal used to generate the comprehensive feature vector D;

[0042] The signal quality verification unit is configured to: when the signal-to-noise ratio of the respiratory vibration signal is lower than a first threshold, or the effective sampling rate of the blood oxygen signal is lower than a second threshold, add a quality attenuation factor to the comprehensive feature vector D, and when the discriminator of the generative adversarial network model library receives the comprehensive feature vector with the quality attenuation factor, correspondingly reduce the confidence level of its output anomaly index.

[0043] Preferably, the discriminator in the generative adversarial network model library includes a feature enhancement layer for performing a nonlinear transformation on the input composite feature vector D before calculating the anomaly index.

[0044] Preferably, the discriminator of the generative adversarial network model adopts a multi-task learning architecture, and its output layer contains two parallel branches: the first branch outputs the data authenticity judgment result, and the second branch outputs the anomaly index.

[0045] Preferably, the system also includes a time-series consistency verification module, used to verify the rationality of the anomaly index in terms of time series; this module calculates the consistency index CI between the current anomaly index and recent historical anomaly indices.

[0046] The technical effects and advantages of this invention are as follows:

[0047] 1. This invention achieves comprehensive and non-intrusive monitoring of users' sleep physiological signals by constructing a multimodal signal acquisition system that combines non-contact and wearable devices. The mattress piezoelectric sensor accurately collects respiratory vibration and body movement signals, while the smart bracelet simultaneously acquires blood oxygen and pulse wave signals. This parallel acquisition mechanism of multi-source heterogeneous signals effectively overcomes the limitations of traditional single-modal monitoring, providing a rich data foundation for subsequent accurate analysis.

[0048] 2. This invention achieves effective fusion and feature enhancement of heterogeneous signals through an innovative multimodal data processing flow. The system uses a first processing method to calculate multimodal correlation, a second processing method to calculate joint temporal variability, and finally generates a comprehensive feature vector through adaptive weighted fusion based on signal-to-noise ratio. This hierarchical and progressive processing mechanism fully explores the complementary information between multimodal signals, significantly improving the quality and robustness of feature representation.

[0049] 3. This invention achieves the dual functions of signal quality verification and anomaly detection simultaneously by designing a multi-task learning architecture for generative adversarial networks. The discriminator includes a feature enhancement layer to improve feature separability, a parallel dual-branch structure to output data authenticity judgment results and anomaly index respectively, and a temporal consistency verification module to ensure the rationality of the output. This integrated design effectively improves the accuracy and reliability of respiratory anomaly detection.

[0050] 4. This invention achieves graded and precise response to respiratory abnormalities by establishing a three-level early warning mechanism based on multi-factor assessment. The system comprehensively considers the abnormality index, signal quality attenuation factor, and temporal consistency index, and sets differentiated early warning thresholds and response strategies. This ensures timely alerts for high-risk situations while avoiding overreaction to transient interference, significantly improving the system's practicality and user experience. Attached Figure Description

[0051] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 This is a diagram of the overall system architecture.

[0053] Figure 2 This is a flowchart of multimodal data processing.

[0054] Figure 3 To generate an architecture diagram for an adversarial network discriminator. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0056] See Figure 1 As shown, the overall architecture of this invention comprises four core modules: a signal acquisition module, a data processing module, a generative adversarial network (GAN) model library, and an early warning module. The signal acquisition module, consisting of a non-contact mattress piezoelectric sensor and a wearable smart bracelet, is responsible for acquiring multimodal physiological signals. The data processing module processes the acquired signals and generates a comprehensive feature vector. The GAN model library receives the feature vector and outputs an anomaly index. The early warning module implements a three-level early warning system based on the anomaly index. These modules are connected sequentially to form a complete monitoring and early warning closed loop.

[0057] The signal acquisition module consists of a non-contact mattress piezoelectric sensor and a wearable smart bracelet, used to collect the following objective physiological signals from the user in parallel: respiratory vibration signals and body movement signals are collected through the mattress piezoelectric sensor; blood oxygen signals, photoplethysmography pulse wave signals and wrist body movement signals are collected through the smart bracelet;

[0058] This embodiment details the hardware selection, deployment method, and data acquisition process of the signal acquisition module. The system employs a non-contact mattress piezoelectric sensor placed under the mattress. This sensor continuously acquires respiratory vibration signals and body movement signals during the user's sleep at a sampling rate of 100Hz. Simultaneously, the user wears a smart bracelet integrating a blood oxygen sensor and a triaxial accelerometer, which acquires blood oxygen saturation signals, photoplethysmography (PPG) signals, and wrist movement signals in parallel at a sampling rate of 50Hz. Both devices establish a communication connection with the data processing module via Bluetooth 5.0, enabling synchronous acquisition and transmission of multimodal signals.

[0059] In a typical application scenario, the system collected the following sample data fragments:

[0060] The mattress piezoelectric sensor outputs data A, in which the amplitude sequence of the breathing vibration signal is [0.12, 0.15, 0.13, 0.11, 0.14] mV, and the corresponding body motion signal energy sequence at that time is [0.02, 0.01, 0.05, 0.03, 0.01] g;

[0061] The smart bracelet synchronously outputs data B, in which the blood oxygen signal value is [96, 95, 97, 96, 95]% and the photoplethysmography (PPG) wave signal amplitude is [210, 205, 215, 208, 212]ADC.

[0062] The energy sequence of the wrist movement signal data C is [0.01, 0.02, 0.01, 0.03, 0.01]g.

[0063] After verification, the raw data is sent to the data processing module for feature extraction, providing a multimodal data foundation for subsequent generative adversarial network models.

[0064] The data processing module, which communicates with the signal acquisition module, is used to process the acquired multimodal objective physiological signals to generate a comprehensive feature vector.

[0065] See Figure 2As shown, the multimodal data processing flow demonstrates the process of generating a comprehensive feature vector from the original signal. The data processing module first defines the input signal as data A, data B, and data C; then, it generates comparison data 1 using a first processing method and comparison data 2 using a second processing method; finally, it integrates the two comparison data to generate a comprehensive feature vector D. This flow achieves feature extraction and fusion of multimodal signals, providing high-quality input features for subsequent anomaly detection.

[0066] This embodiment specifically illustrates the complete process by which the data processing module processes multimodal objective physiological signals to generate comprehensive feature vector data D. The module receives raw data from the signal acquisition module and first classifies the input data according to the definition in claim 2: Data A contains a respiratory vibration signal sequence [0.12, 0.15, 0.13, 0.11, 0.14] mV and a body movement signal sequence [0.02, 0.01, 0.05, 0.03, 0.01] g; Data B contains a blood oxygenation signal sequence [96, 95, 97, 96, 95] % and a photoplethysmography (PPG) pulse wave signal sequence [210, 205, 215, 208, 212] ADC; Data C is a wrist body movement signal sequence [0.01, 0.02, 0.01, 0.03, 0.01] g.

[0067] The data processing module is configured to perform the following operations:

[0068] The respiratory vibration signal and the body movement signal are defined as data A;

[0069] The blood oxygen signal and the photoplethysmography signal are defined as data B;

[0070] The wrist movement signal is defined as data C;

[0071] Data A and data B are processed using the first processing method to generate comparison data 1;

[0072] Data B and data C are processed using the second processing method to generate comparison data 2;

[0073] The comparison data 1 and comparison data 2 are integrated and processed to generate data D as a comprehensive feature vector.

[0074] The first processing method involves calculating the multimodal correlation between data A and data B, specifically including:

[0075] Calculate the first Pearson correlation coefficient between respiratory vibration signal and blood oxygen signal. ;

[0076] Calculate the second Pearson correlation coefficient between the body motion signal and the photoplethysmography (PPG) signal. ;

[0077] Comparison data 1 using the formula:

[0078] Calculated

[0079] in These are the preset weighting coefficients.

[0080] It should be noted that the first Pearson correlation coefficient between the respiratory vibration signal and the blood oxygen signal is calculated first. The mean respiratory vibration signal was 0.13, the mean blood oxygen signal was 95.8, the covariance was calculated to be 0.025, the standard deviation of the respiratory vibration signal was 0.0158, and the standard deviation of the blood oxygen signal was 0.836. .

[0081] Furthermore, the second Pearson correlation coefficient between the body motion signal and the photoplethysmography (PPG) signal was calculated. The mean of the body motion signal is 0.024, the mean of the photoplethysmography (PPG) signal is 210, the calculated covariance is -0.35, the standard deviation of the body motion signal is 0.0179, and the standard deviation of the PPG signal is 3.807. .

[0082] Set preset weight coefficients Comparing data 1 using the formula:

[0083] ,

[0084] Calculated .

[0085] The second processing method is to calculate the joint time-domain variability of data B and data C, specifically obtaining comparison data 2 using the following formula:

[0086] ,

[0087] in, and These are the photoplethysmography (PPG) signal and wrist motion signal values ​​at the i-th sampling point, respectively. Let N be the mean of the product sequence of the two, and N be the number of sampling points within the time window.

[0088] It should be noted that, firstly, the product sequence of the photoplethysmography (PPG) signal and the wrist movement signal is calculated: [210×0.01, 205×0.02, 215×0.01, 208×0.03, 212×0.01] = [2.1, 4.1, 2.15, 6.24, 2.12]. The mean μ of the product sequence is 3.342, and N=5. Then, comparison data 2 is obtained using the formula:

[0089] ,

[0090] Calculated .

[0091] The integration process is an adaptive weighted fusion based on the signal-to-noise ratio, and the formula for generating data D is:

[0092] ,

[0093] Among them, weight The calculation method is as follows:

[0094] , ,

[0095] in, These are the signal-to-noise ratio estimates for data A, data B, and data C within the current time window, respectively.

[0096] The estimated signal-to-noise ratios (SNRs) of data A, B, and C within the current time window are 25 dB, 30 dB, and 20 dB, respectively. Calculate the weights:

[0097] ;

[0098] ;

[0099] Substitute -0.922 from comparison data 1 and 1.209 from comparison data 2 into the formula:

[0100] The final value of the data D, which is used as the comprehensive feature vector, is 0.024. This data D will be directly input into the discriminator of the generative adversarial network model to calculate the anomaly index.

[0101] The data processing module also includes a signal quality verification unit, which is used to evaluate the quality of the original signal used to generate the comprehensive feature vector D;

[0102] The signal quality verification unit is configured to add a quality attenuation factor to the comprehensive feature vector D when the signal-to-noise ratio of the respiratory vibration signal is lower than the first threshold or the effective sampling rate of the blood oxygen signal is lower than the second threshold. When the discriminator of the Generating Adversarial Network Model Library receives the comprehensive feature vector with the quality attenuation factor, it correspondingly reduces the confidence of the anomaly index output by the discriminator.

[0103] This embodiment details the working mechanism of the signal quality verification unit and its impact on the system output. This unit starts immediately after the data processing module generates the comprehensive feature vector D, and performs a quality assessment on the original signal used to generate D. A first threshold of 35dB and a second threshold of 90% are set. The unit first assesses the quality of the respiratory vibration signal in data A, calculating its signal-to-noise ratio to be 25dB, which is lower than the first threshold of 35dB. Simultaneously, it assesses the blood oxygen signal in data B, finding its effective sampling rate within the specified time window to be 85%, lower than the second threshold of 90%. Since both the respiratory vibration signal signal-to-noise ratio and the effective sampling rate of the blood oxygen signal fail to meet the requirements, according to the strictest principle, the signal quality verification unit adds a quality attenuation factor of 0.7 to the comprehensive feature vector D.

[0104] The quality attenuation factor is encapsulated in metadata form along with the value 0.024 of the comprehensive feature vector D, forming a quality-labeled feature vector pair (0.024, 0.7) that is transmitted to the generative adversarial network model library. When the discriminator receives this comprehensive feature vector with the quality attenuation factor, it multiplies the anomaly index by the quality attenuation factor during its internal calculation, thereby reducing the confidence level. For example, if the anomaly index originally output by the discriminator is 0.82, multiplying it by the quality attenuation factor 0.7 adjusts the actual anomaly index used for early warning judgment to 0.574, significantly reducing the possibility of false alarms. Simultaneously, the system records this quality anomaly event for subsequent data analysis and displays a message on the user interface suggesting a change in equipment location due to poor signal quality.

[0105] A generative adversarial network (GAN) model library stores trained GAN models; the discriminator of the GAN model is configured to receive the comprehensive feature vector output by the data processing module and output an abnormality index representing the probability of respiratory abnormalities.

[0106] See Figure 3 As shown, the generative adversarial network (GAN) discriminator architecture comprises a feature enhancement layer, a multi-task learning architecture, and a temporal consistency verification module. The feature enhancement layer performs a nonlinear transformation on the input features; the multi-task learning architecture includes two parallel branches that output the data authenticity judgment result and the anomaly index, respectively; the temporal consistency verification module verifies the temporal rationality of the anomaly index. This architecture ensures the accuracy and reliability of the anomaly index calculation through multiple technical means.

[0107] The discriminator in the Generative Adversarial Network (GAN) model library includes a feature enhancement layer, which performs a nonlinear transformation on the input composite feature vector D before calculating the anomaly index. The transformation formula is as follows:

[0108] ,

[0109] Where W is the weight matrix, b is the bias vector, and ReLU is the activation function; the transformed feature vector D' is then input into the subsequent network layer of the discriminator to calculate the anomaly index.

[0110] The discriminator in the generative adversarial network model employs a multi-task learning architecture. Its output layer contains two parallel branches: the first branch outputs the data authenticity judgment result, and the second branch outputs the anomaly index. The formula for calculating the anomaly index is:

[0111] ,

[0112] Where σ is the Sigmoid activation function. These are the weights and bias parameters of the second branch, respectively.

[0113] This embodiment specifically illustrates the workflow of the discriminator in the Generative Adversarial Network (GAN) model library, which integrates feature enhancement, multi-task learning, and temporal verification.

[0114] The discriminator receives a comprehensive feature vector D from the data processing module, with a value of 0.024. First, this vector is input to the feature enhancement layer for nonlinear transformation. Setting the weight matrix W to [0.8, -0.2] and the bias vector b to 0.1, the transformation process is as follows:

[0115] Then the transformed eigenvector D' is 0.1144.

[0116] Subsequently, according to the multi-task learning architecture, the feature vector D' is simultaneously input to two parallel branches of the discriminator.

[0117] The first branch performs data authenticity determination, and its calculation process is as follows: using a weight vector For [1.5, 0.5] and bias Given -0.2, calculate the linear output:

[0118] ;

[0119] Then, the Sigmoid activation function outputs the true / false probability. Since this probability is greater than 0.5, the first branch determines that the input features are derived from real sleep data.

[0120] The second branch outputs the anomaly index. This branch uses dedicated weights. For [2.0] and bias Given -1.0, calculate the linear output:

[0121] ;

[0122] The anomaly index is then output after passing through the Sigmoid activation function:

[0123] The abnormality index of 0.316 indicates a 31.6% probability of current respiratory abnormality.

[0124] The system also includes a time-series consistency verification module, used to verify the rationality of the anomaly index in time series; this module calculates the consistency index (CI) between the current anomaly index and recent historical anomaly indices:

[0125] ,

[0126] in This is the current abnormal index. The value is the moving average of recent historical anomaly indices, and k is an adjustment parameter. When CI is lower than the preset threshold, the warning module will temporarily suspend issuing warning signals.

[0127] This embodiment illustrates how the module verifies the rationality of the anomaly index. The moving average of recent historical anomaly indices is obtained as 0.28, and the adjustment parameter k is set to 2. The consistency index is calculated:

[0128] .

[0129] Since the CI value of 0.931 is higher than the preset threshold of 0.8, the time series consistency test is passed, and the anomaly index of 0.316 is deemed valid and transmitted to the early warning module for final decision-making.

[0130] The early warning module communicates with the generative adversarial network model library and is used to compare the abnormal index with a preset threshold. When the abnormal index exceeds the preset threshold, it generates and issues a respiratory abnormality early warning signal.

[0131] This embodiment specifically illustrates the three-level early warning mechanism of the early warning module. The module receives an anomaly index of 0.316 from a generative adversarial network model library and combines it with a signal quality attenuation factor of 0.7 and a timing consistency index of 0.931 for comprehensive judgment. The system sets three early warning thresholds: a first-level threshold of 0.3, a second-level threshold of 0.5, and a third-level threshold of 0.7.

[0132] First, the module calculates the effective anomaly index. The original anomaly index of 0.316 is multiplied by the quality decay factor of 0.7 to obtain an adjusted anomaly index of 0.221. Simultaneously, the time series consistency index of 0.931 is verified to be higher than the passing threshold of 0.8, confirming the reasonableness of the data time series.

[0133] Early warning decisions are made based on an adjusted anomaly index of 0.221.

[0134] Level 1 warning: The adjusted abnormal index of 0.221 is lower than the Level 1 warning threshold of 0.3, so the Level 1 warning is not triggered.

[0135] Level 2 warning: The adjusted abnormal index of 0.221 is lower than the Level 2 warning threshold of 0.5, so the Level 2 warning is not triggered.

[0136] Level 3 warning: The adjusted abnormal index of 0.221 is lower than the Level 3 warning threshold of 0.7, so the Level 3 warning is not triggered.

[0137] Since the adjusted anomaly index of 0.221 did not reach any warning level threshold, the warning module remained in normal operation and did not issue any warning signal. Simultaneously, the system recorded the monitoring data, including the original anomaly index of 0.316, the quality decay factor of 0.7, the adjusted anomaly index of 0.221, and the time series consistency index of 0.931, forming a complete monitoring log. The system continues to monitor the anomaly index in subsequent time windows. When the adjusted anomaly index exceeds the corresponding threshold, it will immediately trigger the corresponding level of warning: a level 1 warning prompts the user to check via device indicator lights; a level 2 warning adds a push notification to the mobile app to remind family members to pay attention; and a level 3 warning activates the emergency contact call mechanism.

[0138] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve structures related to the embodiments disclosed in this invention. Other structures can refer to general designs. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other.

[0139] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A sleep quality monitoring and respiratory abnormality early warning system based on generative adversarial networks, characterized in that, include: Signal acquisition module, data processing module, generative adversarial network model library, and early warning module; The signal acquisition module consists of a non-contact mattress piezoelectric sensor and a wearable smart bracelet, and is used to collect the following objective physiological signals of the user in parallel: respiratory vibration signals and body movement signals are collected through the mattress piezoelectric sensor; blood oxygen signals, photoplethysmography pulse wave signals and wrist body movement signals are collected through the smart bracelet; The data processing module is communicatively connected to the signal acquisition module and is used to process the acquired multimodal objective physiological signals to generate a comprehensive feature vector. The generative adversarial network model library stores trained generative adversarial network models; the discriminator of the generative adversarial network model is configured to receive the comprehensive feature vector output by the data processing module and output an abnormality index representing the probability of respiratory abnormality. The early warning module is communicatively connected to the generative adversarial network model library and is used to compare the abnormality index with a preset threshold, and generate and issue a respiratory abnormality early warning signal when the abnormality index exceeds the preset threshold.

2. The sleep quality monitoring and respiratory abnormality early warning system based on generative adversarial networks according to claim 1, characterized in that, The data processing module is configured to perform the following operations: The respiratory vibration signal and body movement signal are defined as data A; The blood oxygen signal and photoplethysmography signal are defined as data B; The wrist movement signal is defined as data C; Data A and data B are processed using the first processing method to generate comparison data 1; Data B and data C are processed using the second processing method to generate comparison data 2; The comparison data 1 and comparison data 2 are integrated and processed to generate data D as the comprehensive feature vector.

3. The sleep quality monitoring and respiratory abnormality early warning system based on generative adversarial networks according to claim 2, characterized in that, The first processing method is to calculate the multimodal correlation between data A and data B, specifically including: Calculate the first Pearson correlation coefficient between respiratory vibration signal and blood oxygen signal. ; Calculate the second Pearson correlation coefficient between the body motion signal and the photoplethysmography (PPG) signal. ; The comparison data 1 was obtained by weighting the first Pearson correlation coefficient and the second Pearson correlation coefficient.

4. The sleep quality monitoring and respiratory abnormality early warning system based on generative adversarial networks according to claim 2, characterized in that, The second processing method involves calculating the joint time-domain variability of data B and data C, specifically obtaining comparison data 2 using the following formula: ， in, and These are the photoplethysmography (PPG) signal and wrist motion signal values ​​at the i-th sampling point, respectively. Let N be the mean of the product sequence of the two, and N be the number of sampling points within the time window.

5. The sleep quality monitoring and respiratory abnormality early warning system based on generative adversarial networks according to claim 2, characterized in that, The integration process is an adaptive weighted fusion based on the signal-to-noise ratio to generate data D.

6. The sleep quality monitoring and respiratory abnormality early warning system based on generative adversarial networks according to claim 1, characterized in that, The data processing module also includes a signal quality verification unit, which is used to evaluate the quality of the original signal used to generate the comprehensive feature vector D. The signal quality verification unit is configured to: when the signal-to-noise ratio of the respiratory vibration signal is lower than a first threshold, or the effective sampling rate of the blood oxygen signal is lower than a second threshold, add a quality attenuation factor to the comprehensive feature vector D, and when the discriminator of the generative adversarial network model library receives the comprehensive feature vector with the quality attenuation factor, correspondingly reduce the confidence level of its output anomaly index.

7. The sleep quality monitoring and respiratory abnormality early warning system based on generative adversarial networks according to claim 1, characterized in that, The discriminator in the generative adversarial network model library includes a feature enhancement layer, which performs a nonlinear transformation on the input composite feature vector D before calculating the anomaly index.

8. The sleep quality monitoring and respiratory abnormality early warning system based on generative adversarial networks according to claim 1, characterized in that, The discriminator of the generative adversarial network model adopts a multi-task learning architecture, and its output layer contains two parallel branches: the first branch outputs the data authenticity judgment result, and the second branch outputs the anomaly index.

9. The sleep quality monitoring and respiratory abnormality early warning system based on generative adversarial networks according to claim 1, characterized in that, The system also includes a time-series consistency verification module, which is used to verify the rationality of the anomaly index in terms of time series; this module calculates the consistency index CI between the current anomaly index and the recent historical anomaly index.

Images