Respiration state real-time monitoring and emergency response system based on AI algorithm

Abstract

The invention relates to the technical field of breathing state monitoring and emergency response systems, and discloses an AI algorithm-based breathing state real-time monitoring and emergency response system, which comprises a data acquisition module, a model training and generalization module, a real-time processing module, a multi-modal information comprehensive evaluation module, a graded early warning and emergency response module and an improvement module. According to the breathing state real-time monitoring and emergency response system based on the AI algorithm, by using the flexible piezoelectric sensor, the band-pass filter and the self-adaptive noise elimination algorithm, it is ensured that high-quality breathing signals are captured, a vivid abnormal breathing mode sample is generated in combination with the generative adversarial network, an intelligent model is constructed by applying the transfer learning technology, and the real-time monitoring and emergency response of the breathing state are achieved. The accuracy and robustness of abnormal breathing pattern recognition are improved, a lightweight neural network architecture and an edge computing concept are adopted, preliminary data analysis and anomaly detection tasks are executed on local equipment, data transmission delay is reduced, and dependence on a cloud server is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of respiratory state monitoring and emergency response systems, and specifically to a real-time respiratory state monitoring and emergency response system based on AI algorithms. Background Art

[0002] A real-time respiratory state monitoring and emergency response system is an integrated health management solution designed to continuously monitor a user's respiratory condition through advanced sensing technologies and artificial intelligence algorithms, and to promptly initiate an emergency response process when an abnormal situation is detected.

[0003] However, traditional methods are unable to effectively remove noise or standardize data when capturing and processing respiratory signals, resulting in reduced accuracy in model training and anomaly detection. Additionally, there are few existing samples of abnormal breathing patterns, making it difficult to cover all possible scenarios and limiting the generalization ability of the model. At the same time, existing systems often use simple binary alert methods and are unable to provide personalized advice or emergency responses based on specific situations. Summary of the Invention

[0004] The purpose of this section is to outline some aspects of the embodiments of the present invention and briefly introduce some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of the application, to avoid obscuring the purpose of this section, the abstract, and the title, but such simplifications or omissions are not intended to limit the scope of the present invention.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] A real-time respiratory state monitoring and emergency response system based on AI algorithms, comprising a data acquisition module, a model training and generalization module, a real-time processing module, a multi-modal information comprehensive evaluation module, a hierarchical early warning and emergency response module, and an improvement module;

[0007] The data acquisition module is used to capture the user's respiratory signal through a sensor, and perform preliminary processing and standardization processing on the raw data to obtain high-quality respiratory data;

[0008] The model training and generalization module constructs a diverse training data set from the obtained high-quality respiratory data, generates samples of abnormal breathing patterns using a generative adversarial network, and constructs an intelligent model using transfer learning techniques;

[0009] The real-time processing module adopts a lightweight neural network architecture, combines the concept of edge computing, performs preliminary data analysis and anomaly detection tasks on a local device, and uses the intelligent model for monitoring to identify abnormal situations;

[0010] The multi-modal information comprehensive evaluation module integrates multiple physiological signal sources, combines abnormal conditions, and constructs a multi-dimensional risk scoring network through machine learning algorithms to calculate the risk score;

[0011] The hierarchical warning and emergency response module sets multiple levels of alarm levels and trigger conditions according to the risk score, and designs an intelligent decision tree to select the action path;

[0012] The improvement module realizes an online learning framework through the action path, enabling the system to learn from new data accumulated in daily operations and continuously optimize.

[0013] As a further solution of the present invention: The method for capturing the user's breathing signal through a sensor, and performing preliminary cleaning and standardization processing on the original data to obtain high-quality breathing data, the specific steps are as follows:

[0014] Use a flexible piezoelectric sensor to capture the user's breathing movement;

[0015] Apply a band-pass filter to remove unnecessary frequency components in the user's breathing movement, and use an adaptive noise cancellation algorithm to reduce background noise;

[0016] Normalize the signal so that all data points fall within a fixed range;

[0017] Use the weighted average method to calculate the final comprehensive signal in combination with the different characteristics of each sensor to obtain high-quality breathing data.

[0018] As a further solution of the present invention: The method for constructing a diverse training dataset through the obtained high-quality breathing data and generating abnormal breathing pattern samples using a generative adversarial network, the specific steps are as follows:

[0019] Collect data samples from a wide range of people, including normal and abnormal breathing patterns of individuals of different ages, genders, and health conditions in various environments, and use a sampling method based on information entropy to calculate the number of breathing pattern categories;

[0020] Set the generator G(z; θ g ) and the discriminator D(x; θ d ), where z is a random noise vector from the prior distribution p z (z), θ g and θ d are the parameters of the generator and the discriminator respectively. The generator creates a realistic abnormal breathing signal x' = G(z; θ g ), while the discriminator tries to distinguish between real samples x and generated samples x';

[0021] Generate abnormal breathing pattern samples, and the expression is:

[0022]

[0023] Among them, z is random noise drawn from a predefined probability distribution, is the optimal generator parameter obtained through the adversarial training process, enabling the generator to generate abnormal breathing pattern samples as realistic as possible, and W is the abnormal breathing pattern sample.

[0024] As a further solution of the present invention: The intelligent model is constructed by applying transfer learning technology, and the specific steps are as follows:

[0025] Set the intelligent model Y by combining the source domain and the target domain, and optimize it using the transfer learning loss function. The expression is:

[0026] Y(x; θ * ) = f base (x; θ b ) + β · f target (W; θ t );

[0027] L t = α(L S (θ b ) + L w ) + (1 - α)L T (θ t );

[0028] Among them, x is the input breathing signal, f base (x; θ b ) is the basic model pre-trained on a large general dataset, with parameters θ b , f target (W; θ t ) is the part specifically fine-tuned for the target domain, with parameters θ t , β is a hyperparameter controlling the balance between the basic model and the target domain part, W is the abnormal breathing pattern sample generated by the adversarial generation network, L S (θ b ) is the loss on the source domain, used to evaluate the performance of the basic model f base on the general dataset, L w is the adversarial generation network loss weighted by information entropy, L T (θ t ) is the loss on the target domain, used to evaluate the performance of the model f target in a specific user or niche group, and α is a hyperparameter controlling the balance between the two domains.

[0029] As a further solution of the present invention: By adopting a lightweight neural network architecture and combining the concept of edge computing, preliminary data analysis and anomaly detection tasks are performed on local devices. The specific steps are as follows:

[0030] Select MobileNetV3 as the lightweight neural network architecture, deploy the inference process to the user's wearable device or mobile device, and utilize the concept of edge computing to reduce the dependence on the cloud server;

[0031] When the breathing signal enters the system, preliminary data analysis is first performed, and feature extraction is carried out. The expression is:

[0032]

[0033] Among them, A i (t) represents the amplitude value of the i-th sensor at time t, W i is the weight determined according to the sensor reliability and signal strength, N is the number of sensors participating in the fusion, P(f) is the power spectral density function, representing the energy distribution at frequency f, and F(t) is the feature extraction result.

[0034] As a further solution of the present invention: Use an intelligent model for monitoring and identifying abnormal situations. The specific steps are as follows:

[0035] Based on the feature extraction result F(t), use the pre-trained intelligent model Y to perform anomaly detection on the input breathing signal and output the anomaly detection result. The expression is:

[0036]

[0037] Among them, S a (x; θ * ) is the anomaly score, indicating the possibility that the input breathing signal x belongs to the abnormal pattern, γ(Y(x; θ * ) is the output of the intelligent model, with the parameter θ * , γ is the sensitivity coefficient, and T is the preset threshold used to distinguish normal and abnormal states.

[0038] As a further solution of the present invention: Integrate multiple physiological signal sources, combine abnormal situations, and construct a multi-dimensional risk scoring network through machine learning algorithms to calculate the risk score. The specific steps are as follows:

[0039] Take the anomaly detection result S a (x; θ * ) provided by the real-time processing module as an additional input feature and add it to the comprehensive feature vector x to form an extended feature vector x′. The expression is:

[0040] X′ = [X, Sa (x; θ * )];

[0041] Based on the extended feature vector X′, design a multi-dimensional risk scoring network R(X′; θ r );

[0042] Introduce an attention mechanism, and the final risk score R is expressed as:

[0043] R(X′; θ r ) = σ(W 2 ·ReLU(W 1 ·X′ + b 1 ) + b 2 );

[0044] Among them, σ is the sigmoid activation function, W 1 and W 2 are weight matrices, b 1 and b 2 are bias terms, ReLU is the rectified linear unit activation function, used to introduce non-linearity;

[0045] Obtain the final risk score R.

[0046] As a further solution of the present invention: According to the risk score, set multiple levels of alarm levels and trigger conditions, and the specific steps are as follows:

[0047] Divide the risk score R into multiple alarm levels according to its value range [0, 1], specifically as follows:

[0048] Low risk: When R ∈ [0, 0.3), it means that the user's state is basically normal;

[0049] Medium risk: When R ∈ [0.3, 0.7), it means that the user has potential health problems, and it is recommended to further observe or take preventive measures;

[0050] High risk: When R ∈ [0.7, 1], it means that the user is in an emergency and immediate medical intervention is required.

[0051] As a further solution of the present invention: Each alarm level corresponds to different trigger conditions, and the conditions are not only based on the risk score R, but also combined with the change trend of other physiological signals and historical data, specifically as follows:

[0052] Low risk: If the R values for three consecutive times are all lower than 0.3, the system sends a prompt message, suggesting that the user maintain the current behavior;

[0053] Medium risk: If R continuously stays between 0.3 and 0.7 for more than 10 minutes, the system sends a notice to the user's contacts or family members and recommends that the user perform a self-examination;

[0054] High risk: If R exceeds 0.7 and the duration exceeds 5 minutes, the system automatically triggers an emergency response mechanism, including calling the emergency service, notifying the medical staff, and providing real-time location information.

[0055] As a further solution of the present invention: Through the action path, an online learning framework is realized, enabling the system to learn from new data accumulated in daily operations and continuously optimize. The specific steps are as follows:

[0056] The system continuously collects the physiological signals and behavioral data of the user in different scenarios;

[0057] Each time the system triggers an alarm or executes an emergency response, it will record the relevant anomaly detection results S in detail a and the risk score R;

[0058] An incremental data processing method is adopted to maintain the real-time performance and efficiency of the system. The newly collected data will be regularly summarized and merged with the historical data to form a dynamic data set;

[0059] Based on the dynamic data set, an online learning framework is designed to allow the model to be gradually optimized without interrupting the service. The expression is:

[0060]

[0061] where, are the risk score network parameters at the t-th iteration, η is the learning rate, represents the gradient of the parameter θ r L is the loss function, and y is the actual label.

[0062] Compared with the prior art, the beneficial effects of the present invention are:

[0063] By using flexible piezoelectric sensors, band-pass filters, and adaptive noise cancellation algorithms, high-quality respiratory signals are ensured to be captured. Combining generative adversarial networks to generate realistic samples of abnormal breathing patterns and applying transfer learning techniques to build intelligent models improve the accuracy and robustness of abnormal breathing pattern recognition. Adopting a lightweight neural network architecture and edge computing concept to perform preliminary data analysis and anomaly detection tasks on local devices not only reduces data transmission latency but also decreases dependence on cloud servers, enhancing the system's response speed and privacy protection level. Integrating multiple physiological signal sources and constructing a multi-dimensional risk scoring network through machine learning algorithms, and introducing an attention mechanism enables the model to automatically learn the importance of different features, thus more accurately evaluating the user's health status. Setting multi-level alarm levels and trigger conditions based on the risk score and designing an intelligent decision tree to select action paths, the hierarchical early warning mechanism can provide personalized suggestions or emergency responses according to the user's specific situation, effectively improving the ability to respond to sudden health events. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 Schematic diagram of a real-time monitoring and emergency response system for respiratory status based on AI algorithms. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0065] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following detailed description of the specific embodiments of the present invention will be given with reference to the accompanying drawings of the specification.

[0066] In the following description, many specific details are set forth to facilitate a thorough understanding of the present invention. However, the present invention may be implemented in other ways different from those described herein. Those skilled in the art can make similar generalizations without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0067] Secondly, the so-called "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation manner of the present invention. The appearances of "in one embodiment" in different places in this specification do not all refer to the same embodiment, nor are they separate or alternative embodiments that exclude each other.

[0068] Embodiment 1

[0069] Please refer to Figure 1 , which is the first embodiment of the present invention. This embodiment provides a real-time monitoring and emergency response system for respiratory status based on AI algorithms, including:

[0070] A data acquisition module, a model training and generalization module, a real-time processing module, a multi-modal information comprehensive evaluation module, a hierarchical early warning and emergency response module, and an improvement module;

[0071] A data acquisition module, which is used to capture the user's breathing signal through a sensor, and perform preliminary cleaning and standardization processing on the original data to obtain high-quality breathing data;

[0072] Use a flexible piezoelectric sensor to capture the user's breathing movement;

[0073] Apply a band-pass filter to remove unwanted frequency components in the user's breathing movement, and use an adaptive noise cancellation algorithm to reduce background noise;

[0074] Normalize the signal so that all data points fall within a fixed range;

[0075] Use the weighted average method to calculate the final comprehensive signal by combining the different characteristics of each sensor to obtain high-quality breathing data.

[0076] A model training and generalization module, which constructs a diverse training data set through the obtained high-quality breathing data, uses a generative adversarial network to generate abnormal breathing pattern samples, and applies transfer learning technology to construct an intelligent model;

[0077] Collect data samples from a wide range of people, including normal and abnormal breathing patterns of individuals of different ages, genders, and health conditions in various environments, and use a sampling method based on information entropy to calculate the number of breathing pattern categories;

[0078] Set the generator G(z; θ g ) and the discriminator D(x; θ d ), where z is a random noise vector from the prior distribution p z (z), θ g and θ d are the parameters of the generator and the discriminator respectively. The generator creates a realistic abnormal breathing signal x' = G(z; θ g ), while the discriminator tries to distinguish between the real sample x and the generated sample x';

[0079] Generate abnormal breathing pattern samples, and the expression is:

[0080]

[0081] where z is a random noise drawn from a predefined probability distribution, is the optimal generator parameter obtained through the adversarial training process, enabling the generator to generate as realistic abnormal breathing pattern samples as possible, and W is the abnormal breathing pattern sample;

[0082] Combine the source domain and the target domain to set the intelligent model Y, and optimize it using the transfer learning loss function. The expression is:

[0083] Y(x; θ* ) = f base (x; θ b ) + β·f target (W; θ t );

[0084] L t = α(L S (θ b ) + L w )(1 - α)L T (θ t );

[0085] Among them, x is the input respiratory signal, f base (x; θ b ) is the basic model pre-trained on a large general dataset with parameters θ b , f target (W; θ t ) is the part specifically fine-tuned for the target domain with parameters θ t , β is a hyperparameter that controls the balance between the basic model and the target domain part, W is the abnormal respiratory pattern sample generated by the adversarial generation network, L S (θ b ) is the loss on the source domain, used to evaluate the performance of the basic model f base on the general dataset, L w is the adversarial generation network loss weighted by information entropy, L T (θ t ) is the loss on the target domain, used to evaluate the performance of the model f target in a specific user or niche group, and α is a hyperparameter that controls the balance between the two domains.

[0086] The real-time processing module adopts a lightweight neural network architecture, combines the concept of edge computing, performs preliminary data analysis and anomaly detection tasks on local devices, and uses an intelligent model for monitoring to identify abnormal situations;

[0087] Select MobileNetV3 as the lightweight neural network architecture, deploy the inference process to the user's wearable device or mobile device, and use the concept of edge computing to reduce the dependence on the cloud server;

[0088] When the respiratory signal enters the system, preliminary data analysis is first performed, and feature extraction is carried out. The expression is:

[0089]

[0090] Among them, A i (t) represents the amplitude value of the i-th sensor at time t, W iThe weight is determined based on sensor reliability and signal strength, N is the number of sensors participating in the fusion, P(f) is the power spectral density function, representing the energy distribution at frequency f, and F(t) is the feature extraction result;

[0091] Based on the feature extraction result F(t), the pre-trained intelligent model Y is used to perform anomaly detection on the input respiratory signal and output the anomaly detection result. The expression is:

[0092]

[0093] Among them, S a (x; θ * ) is the anomaly score, indicating the possibility that the input respiratory signal x belongs to the abnormal pattern. γ(Y(x; θ * ) is the output of the intelligent model, with the parameter θ * , γ is the sensitivity coefficient, and T is the preset threshold used to distinguish normal and abnormal states.

[0094] The multi-modal information comprehensive evaluation module integrates multiple physiological signal sources, combines the abnormal conditions, and constructs a multi-dimensional risk scoring network through machine learning algorithms to calculate the risk score;

[0095] The anomaly detection result S a (x; θ * ) provided by the real-time processing module is added as an additional input feature to the comprehensive feature vector x to form the extended feature vector x'. The expression is:

[0096] X' = [X, S a (x; θ * )];

[0097] Based on the extended feature vector X', a multi-dimensional risk scoring network R(X'; θ r ) is designed;

[0098] The attention mechanism is introduced, and the final risk score R is expressed as:

[0099] R(X'; θ r ) = σ(W 2 ·ReLU(W 1 ·X' + b 1 ) + b 2 );

[0100] Among them, σ is the sigmoid activation function, W 1 and W 2 are weight matrices, b 1 and b 2 are bias terms, and ReLU is the rectified linear unit activation function used to introduce non-linearity;

[0101] Obtain the final risk score R.

[0102] The grading warning and emergency response module sets multiple levels of alarm levels and trigger conditions according to the risk score, and designs an intelligent decision tree to select the action path;

[0103] Divide the risk score R into multiple alarm levels according to its value range [0, 1] as follows:

[0104] Low risk: When R ∈ [0, 0.3), it indicates that the user's state is basically normal;

[0105] Medium risk: When R ∈ [0.3, 0.7), it indicates that the user has potential health problems, and it is recommended to further observe or take preventive measures;

[0106] High risk: When R ∈ [0.7, 1], it indicates that the user is in an emergency and immediate medical intervention is required;

[0107] Each alarm level corresponds to different trigger conditions, which are based not only on the risk score R but also on the change trends of other physiological signals and historical data, as follows:

[0108] Low risk: If the R values for three consecutive times are all lower than 0.3, the system sends a prompt message, recommending that the user maintain the current behavior;

[0109] Medium risk: If R continuously stays between 0.3 and 0.7 for more than 10 minutes, the system sends a notice to the user's contacts or family members and recommends that the user conduct a self-examination;

[0110] High risk: If R exceeds 0.7 and the duration exceeds 5 minutes, the system automatically triggers an emergency response mechanism, including calling the emergency service, notifying the medical staff, and providing real-time location information.

[0111] The improvement module realizes an online learning framework through the action path, enabling the system to learn from new data accumulated in daily operations and continuously optimize;

[0112] The system continuously collects the user's physiological signals and behavior data in different scenarios;

[0113] Each time the system triggers an alarm or executes an emergency response, it will record the relevant anomaly detection results S in detail a and the risk score R;

[0114] Adopt an incremental data processing method to maintain the real-time performance and efficiency of the system. The newly collected data will be regularly summarized and merged with the historical data to form a dynamic data set;

[0115] Based on a dynamic dataset, an online learning framework is designed to allow the model to be gradually optimized without interrupting the service, with the expression:

[0116]

[0117] where are the risk scoring network parameters at the t-th iteration, η is the learning rate, represents the gradient of the parameter θ r with respect to, L is the loss function, and y is the actual label.

[0118] In summary, by using flexible piezoelectric sensors, band-pass filters, and adaptive noise cancellation algorithms, high-quality respiratory signals are ensured to be captured. Combining with a generative adversarial network to generate realistic abnormal breathing pattern samples, and applying transfer learning techniques to build an intelligent model, the accuracy and robustness of abnormal breathing pattern recognition are improved. By adopting a lightweight neural network architecture and edge computing concept, preliminary data analysis and anomaly detection tasks are performed on local devices, which not only reduces data transmission latency but also decreases the dependence on cloud servers, enhancing the system's response speed and privacy protection level. Integrating multiple physiological signal sources and constructing a multi-dimensional risk scoring network through machine learning algorithms, and introducing an attention mechanism enables the model to automatically learn the importance of different features, thereby more accurately evaluating the user's health status. Setting multi-level alarm levels and trigger conditions according to the risk score, and designing an intelligent decision tree to select the action path, the hierarchical early warning mechanism can provide personalized suggestions or emergency responses according to the specific situation of the user, effectively improving the ability to respond to sudden health events.

[0119] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered within the scope of the claims of the present invention.

Claims

1. A real-time respiratory status monitoring and emergency response system based on AI algorithm, characterized by: include: Data acquisition module, model training and generalization module, real-time processing module, multimodal information comprehensive evaluation module, graded warning and emergency response module and improvement module; The data acquisition module is used to capture the user's breathing signal through a sensor, and perform preliminary processing and standardization on the raw data to obtain high-quality breathing data; The model training and generalization module constructs a diversified training data set through the obtained high-quality breathing data, generates abnormal breathing pattern samples using a generative adversarial network, and constructs an intelligent model using transfer learning technology; The real-time processing module adopts a lightweight neural network architecture and combines the concept of edge computing to perform preliminary data analysis and anomaly detection tasks on local devices, using intelligent models for monitoring and identifying abnormal situations; The multimodal information comprehensive evaluation module integrates multiple physiological signal sources, combines abnormal conditions, and constructs a multi-dimensional risk scoring network through a machine learning algorithm to calculate the risk score; The hierarchical warning and emergency response module sets multiple levels of alarm levels and trigger conditions according to risk scores, and designs an intelligent decision tree to select an action path; The improvement module implements an online learning framework through an action path, enabling the system to learn from new data accumulated in daily operations and continuously optimize.

2. The AI ​​algorithm-based respiratory status real-time monitoring and emergency response system according to claim 1, characterized in that: The method is used to capture the user's breathing signal through the sensor, and perform preliminary cleaning and standardization processing on the raw data to obtain high-quality breathing data. The specific steps are as follows: Use flexible piezoelectric sensors to capture the user's breathing movements; A bandpass filter is applied to remove unwanted frequency components from the user's breathing movements, and an adaptive noise cancellation algorithm is used to reduce background noise; Normalize the signal so that all data points fall within a fixed range; The weighted average method is used to combine the different characteristics of each sensor to calculate the final comprehensive signal and obtain high-quality respiratory data.

3. The AI ​​algorithm-based respiratory status real-time monitoring and emergency response system according to claim 1, characterized in that: The method constructs a diversified training data set through the obtained high-quality breathing data, and uses the adversarial generative network to generate abnormal breathing pattern samples, and the specific steps are as follows: Collect data samples from a wide range of people, including normal and abnormal breathing patterns of individuals of different ages, genders, and health conditions in various environments, and use a sampling method based on information entropy to calculate the number of breathing pattern categories; Set the generator G(z;θ g ) and the discriminator D(x;θ d ), where z is from the prior distribution p z (z) is a random noise vector, θ g and θ d are the parameters of the generator and discriminator respectively. The generator creates a realistic abnormal breathing signal x′=G(z;θ g ), while the discriminator tries to distinguish between real samples x and generated samples x′; Generate abnormal breathing pattern samples, the expression is: where z is random noise drawn from a predefined probability distribution, is the optimal generator parameter obtained through the adversarial training process, so that the generator can generate abnormal breathing pattern samples that are as realistic as possible, and W is the abnormal breathing pattern sample.

4. The AI ​​algorithm-based respiratory status real-time monitoring and emergency response system according to claim 3 is characterized in that: The application of transfer learning technology to build an intelligent model includes the following specific steps: Combine the source domain and the target domain to set the intelligent model Y, and use the transfer learning loss function for optimization, which is expressed as: Y(x;θ * )=f base (x;θ b )+β·f target (W;θ t ); L t =a(L S (i b )+L w )+(1-α)L T (i t ); Where x is the input breathing signal, f base (x;θ b ) is a base model pre-trained on a large general dataset with parameters θ b , f target (W;θ t ) is the part specifically fine-tuned for the target domain, with the parameter θ t , β is a hyperparameter that controls the balance between the base model and the target domain part, W is the abnormal breathing pattern sample generated by the adversarial generative network, and L S (θ b ) is the loss on the source domain, used to evaluate the base model f base Performance on common datasets, L w is the information entropy weighted adversarial network loss, L T (θ t ) is the loss on the target domain, used to evaluate the model f target Performance among specific users or niche groups, α is a hyperparameter that controls the balance between the two areas.

5. The AI ​​algorithm-based respiratory status real-time monitoring and emergency response system according to claim 1, characterized in that: The lightweight neural network architecture is combined with the edge computing concept to perform preliminary data analysis and anomaly detection tasks on the local device. The specific steps are as follows: MobileNetV3 is selected as the lightweight neural network architecture, and the reasoning process is deployed on the user's wearable device or mobile device, using the concept of edge computing to reduce dependence on cloud servers; When the breathing signal enters the system, a preliminary data analysis is first performed and feature extraction is performed. The expression is: Among them, A i (t) represents the amplitude value of the i-th sensor at time t, W i is the weight determined according to the reliability and signal strength of the sensor, N is the number of sensors involved in the fusion, P(f) is the power spectral density function, representing the energy distribution at frequency f, and F(t) is the feature extraction result.

6. The AI ​​algorithm-based real-time respiratory status monitoring and emergency response system according to claim 5 is characterized in that: The specific steps of using intelligent models to monitor and identify abnormal situations are as follows: Based on the feature extraction result F(t), the pre-trained intelligent model Y is used to detect anomalies on the input respiratory signal and output the anomaly detection result, which is expressed as: Among them, S a (x;θ * ) is the anomaly score, which indicates the possibility that the input respiratory signal x belongs to an abnormal mode, γ(Y(x;θ * ) is the output of the intelligent model, and the parameter is θ * , γ is the sensitivity coefficient, and T is the preset threshold used to distinguish between normal and abnormal states.

7. The AI ​​algorithm-based respiratory status real-time monitoring and emergency response system according to claim 1, characterized in that: The method integrates multiple physiological signal sources, combines abnormal conditions, constructs a multi-dimensional risk scoring network through a machine learning algorithm, and calculates the risk score. The specific steps are: The anomaly detection result S provided by the real-time processing module a (x;θ * ) is added as an additional input feature to the comprehensive feature vector x to form the expanded feature vector x′, which is expressed as: X′=[X,S a (x;θ * )]; Based on the expanded feature vector X′, a multi-dimensional risk scoring network R(X′; θ r ); By introducing the attention mechanism, the final risk score R is expressed as: R(X′;θ r )=σ(W2·ReLU(W1·X′+b1)+b2); Among them, σ is the sigmoid activation function, W1 and W2 are weight matrices, b1 and b2 are bias terms, and ReLU is the rectified linear unit activation function, which is used to introduce nonlinearity; The final risk score R is obtained.

8. The AI ​​algorithm-based respiratory status real-time monitoring and emergency response system according to claim 1, characterized in that: According to the risk score, multiple levels of alarm levels and trigger conditions are set. The specific steps are as follows: According to the value range [0,1] of the risk score R, it is divided into multiple alert levels, as follows: Low risk: When R∈[0,0.3), it means that the user status is basically normal; Medium risk: When R∈[0.3,0.7), it means that the user has potential health problems and further observation or preventive measures are recommended; High risk: When R∈[0.7,1], it indicates that the user is in an emergency situation and requires immediate medical intervention.

9. The AI ​​algorithm-based real-time respiratory status monitoring and emergency response system according to claim 8, characterized in that: Each alarm level corresponds to different trigger conditions, which are not only based on the risk score R, but also combined with the change trend and historical data of other physiological signals, as follows: Low risk: If the R value is less than 0.3 for three consecutive times, the system will issue a prompt message and recommend that the user maintain the current behavior; Medium risk: If R remains between 0.3 and 0.7 for more than 10 minutes, the system sends a notification to the user’s contacts or family members and advises the user to conduct a self-examination; High risk: If R exceeds 0.7 and lasts for more than 5 minutes, the system automatically triggers the emergency response mechanism, including calling emergency services, notifying medical staff and providing real-time location information.

10. The AI ​​algorithm-based respiratory status real-time monitoring and emergency response system according to claim 1, characterized in that: The action path is used to implement an online learning framework, which enables the system to learn and continuously optimize from new data accumulated in daily operations. The specific steps are: The system continuously collects the user's physiological signals and behavioral data in different scenarios; Each time the system triggers an alarm or performs an emergency response, the relevant anomaly detection results S are recorded in detail. a and risk score R; Incremental data processing methods are used to maintain the real-time and efficiency of the system. Newly collected data will be regularly summarized and merged with historical data to form a dynamic data set; Based on the dynamic data set, an online learning framework is designed to allow the model to be gradually optimized without interrupting the service. The expression is: in, is the risk scoring network parameter at the tth iteration, η is the learning rate, Represents the parameter θ r The gradient of , L is the loss function, and y is the actual label.

Images