Breathing anomaly detection method and system based on gas monitoring

Abstract

The invention discloses a breath anomaly detection method and system based on gas monitoring, and relates to the technical field of breath anomaly detection.The breath anomaly detection method comprises the steps that mouth and nose gas signals in the breath process of a user are collected and preprocessed, and gas component separation is conducted through a microfluidic gas separation module; and performing differential analysis on separated gas components, calculating fluctuation of nasal cavity and oral cavity gas components, performing breathing mode recognition through 1D SNN, and establishing a personalized breathing mode library in combination with long-term health data of a user. Compared with the prior art, the method has the advantages that more accurate breathing mode recognition and anomaly detection are realized by combining differential analysis of multi-dimensional gas components and a deep learning model with a personalized breathing mode library, so that a more intelligent health management scheme is provided, and particularly, the method has remarkable advantages in coping with complex personalized differences and long-term health trends.

Description

Technical Field

[0001] This invention relates to the field of respiratory abnormality detection technology, and in particular to a respiratory abnormality detection method and system based on gas monitoring. Background Technology

[0002] With the development of respiratory health monitoring technology, an increasing number of gas monitoring-based methods are being applied to the early diagnosis of diseases and health management. Gas monitoring technology provides real-time physiological feedback by capturing and analyzing changes in the gaseous components of human respiration (such as oxygen, carbon dioxide, and volatile organic compounds). For example, changes in exhaled carbon dioxide concentration can reflect lung ventilation, while changes in volatile organic compounds are closely related to metabolism, health status, or disease risk (such as diabetes and asthma). In recent years, advancements in sensor technology, microfluidics, and machine learning algorithms have significantly improved the sensitivity and accuracy of gas sensors, providing more effective tools for respiratory health monitoring, particularly in respiratory pattern recognition, anomaly detection, and disease prediction.

[0003] However, existing gas monitoring-based respiratory health monitoring technologies still face several challenges, particularly in real-time monitoring and accurate diagnosis. Traditional methods often rely on the analysis of a single gas component (such as carbon dioxide or oxygen concentration), making it difficult to simultaneously consider changes in multiple gas components and their interactions, thus limiting the comprehensiveness and accuracy of the detection. Furthermore, existing pattern recognition methods largely depend on statistical or simple machine learning models, which cannot effectively handle and identify complex breathing patterns, especially limiting personalized breathing pattern recognition and anomaly detection. In addition, existing breathing pattern libraries also face insufficient handling of individual differences when matching long-term health data with real-time signals, thus failing to provide customized health management recommendations based on each user's unique physiological characteristics. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a respiratory anomaly detection method and system based on gas monitoring, which solves the shortcomings of existing technologies in real-time, accurate anomaly detection and personalized health management plan generation.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for detecting abnormal breathing based on gas monitoring, comprising, The system collects and preprocesses the gas signals from the user's mouth and nose during breathing, and then separates the gas components using a microfluidic gas separation module. Differential analysis was performed on the separated gas components to calculate the fluctuations in nasal and oral gas components, and breathing patterns were identified using 1DSNN. At the same time, a personalized breathing pattern library was established by combining the user's long-term health data. The system detects anomalies in real-time gas signals and compares them with a personalized breathing pattern library. If an anomaly is detected, an early warning notification is issued immediately, the breathing assistance mode is activated, and personalized health recommendations are generated.

[0007] As a preferred embodiment of the respiratory abnormality detection method based on gas monitoring described in this invention, wherein: the respiratory pattern recognition through 1D SNN and the establishment of a personalized respiratory pattern library in combination with the user's long-term health data refer to the construction of a pattern recognition model based on the 1D SNN model. Real-time differential gas fluctuation sequence segments are captured by a sliding window and fed into a pattern recognition model for processing to obtain the corresponding breathing patterns. Collect users' historical breathing-related data, standardize it, perform pattern recognition, and store it in a personalized database. Calculate the similarity between the data and the existing embedded vectors in the database. If the similarity is higher than a set threshold A, it is determined to be a known breathing pattern; otherwise, it is determined to be a new pattern and enters the labeling queue.

[0008] As a preferred embodiment of the gas monitoring-based respiratory anomaly detection method of the present invention, the step of detecting anomalies in real-time gas signals and comparing them with a personalized breathing pattern library includes: Historical gas signals were collected to generate historical time-series data. A sliding window method was used to segment the time-series data into data segments, with each time window representing an input sample. All sample segments were standardized, and all standardized time segments were stored as a training dataset. Maximum overlap discrete wavelet transform was performed on each input time segment, and the Coiflet 5 wavelet function was selected for signal decomposition. Wavelet decomposition levels were set, and wavelet coefficients at each level were obtained. The wavelet coefficients of each layer are concatenated into an enhanced feature vector X. t ; An encoder network is constructed using LSTM layers to enhance the input feature vector X. t Extract temporal features and generate a low-dimensional latent space representation Z. t The LSTM decoder maps the low-dimensional latent space representation generated by the encoder back to the original signal space, reconstructing an approximation of the input signal. ; By optimizing the loss function To minimize the reconstruction error, the LSTM-Autoencoder model can reconstruct a result that is as close as possible to the input signal. By continuously adjusting the network parameters, the model will learn to extract effective temporal features from the input signal and map them back to the original signal space. The Adam optimizer is used to update the network weights. An early stopping strategy is adopted, and training is automatically stopped when the loss on the validation set no longer improves, resulting in a well-trained LSTM-Autoencoder model. The real-time enhanced feature vectors are input into the trained LSTM-Autoencoder model to obtain the reconstructed signal. The original input signal and the reconstructed signal output by the model are compared, and the reconstruction error is calculated using Euclidean distance. A dynamic anomaly detection threshold θ is set. If the reconstruction error at the current time exceeds the threshold θ, the sample is judged as an anomaly. Based on the reconstruction error judgment, the embedding features Z of the encoder output are processed using a One-Class SVM. t Boundary strengthening is performed to further confirm anomalous samples, and Z is... t As input, a personalized breathing pattern library is queried for comparison. During the comparison process, the similarity between the current vector and each reference vector in the library is calculated one by one using Euclidean distance. If the similarity between the current embedded vector and the pattern in the library exceeds a preset threshold, the current signal is determined to be a known normal behavior pattern; otherwise, it is marked as an unknown behavior pattern. The three judgment results are combined: whether the reconstruction error exceeds the threshold, whether the One-Class SVM judgment is out of bounds, and whether it matches the personalized pattern library. The combination of these three indicators constitutes the basis of the joint anomaly judgment mechanism and divides the samples into different levels, including mild anomaly, moderate anomaly, severe anomaly, and new pattern to be confirmed.

[0009] As a preferred embodiment of the respiratory anomaly detection method based on gas monitoring described in this invention, the step of activating the respiratory support mode and generating personalized health recommendations includes: When the joint anomaly determination mechanism outputs a severe anomaly or a new mode to be confirmed and the anomaly type in the personalized breathing mode library is a severe respiratory anomaly, the respiratory status is determined to be critical and the "respiratory support mode" is immediately triggered. Collect and integrate user health data and environmental information and preprocess them. Continuous health data is divided into small segments according to time windows. Statistical features within each time window are calculated and combined into a feature vector. Multiple segment-level feature vectors are concatenated to form a period-level health status representation and generate a period-level health status vector. The weight of each segment or period feature is calculated through an attention mechanism to obtain a weighted health status vector and generate a preliminary health management plan. The Hybrid Optimization Algorithm WCSO is employed, combining the global search capability of the water cycle algorithm with the local refinement capability of the competitive swarm optimization algorithm to iteratively optimize the health management plan. Based on the optimization plan output by WCA and CSO, and combined with the user's health status vector, different weights are assigned to each health suggestion. Personalized health management suggestions, including short-term and long-term suggestions, are generated based on the weighted plan. The generated personalized health suggestions are pushed to the user, and the health management plan is dynamically adjusted based on user feedback.

[0010] As a preferred embodiment of the respiratory abnormality detection method based on gas monitoring described in this invention, the step of performing differential analysis on the separated gas components and calculating the fluctuation of nasal and oral gas components refers to performing differential analysis on the nasal and oral gas component signals, calculating the component difference between the two at the same time point, obtaining the fluctuation of gas components, and using the root mean square value to calculate the amplitude of the differential gas component fluctuation signal to quantify the degree of fluctuation.

[0011] As a preferred embodiment of the respiratory anomaly detection method based on gas monitoring described in this invention, the step of collecting and preprocessing the oral and nasal gas signals during the user's breathing process refers to collecting raw data on the gas concentration in the nasal cavity and oral cavity. Each set of data includes real-time concentration readings of carbon dioxide, oxygen, and volatile organic compounds in the nasal cavity and oral cavity. A high-pass filter is used to remove high-frequency noise from the gas signals, and a fast Fourier transform is used to perform frequency domain analysis on the noise-removed data.

[0012] As a preferred embodiment of the respiratory abnormality detection method based on gas monitoring described in this invention, the gas component separation by the microfluidic gas separation module refers to locating the main frequency peak of exhalation according to the fast Fourier transform result, and introducing the nasal and oral gases in the current channel into the microfluidic separation chip respectively. After initial flow regulation of the gas to ensure a stable laminar flow state, the gas flows into the separation channel. Guided by multiple transverse micro-branch channels, the components undergo diffusion separation to obtain the separated gas components.

[0013] Secondly, the present invention provides a respiratory anomaly detection system based on gas monitoring, comprising, The gas signal acquisition module is used to acquire the user's oral and nasal gas signals, preprocess them, remove noise, and perform frequency domain analysis. A microfluidic gas separation module is used to separate the gas components in the oral cavity and nasal cavity using microfluidic technology; The differential analysis module is used to calculate the fluctuations in the gas composition of the nasal cavity and oral cavity, and to quantify the differences in gas composition. The breathing pattern recognition module is used to recognize breathing patterns using 1D SNN and to build a personalized breathing pattern library by combining user health data. The health advice generation module is used to detect anomalies in real-time gas signals and compare them with a personalized breathing pattern library. Based on the anomaly detection results, it activates the breathing assistance mode and generates personalized health advice.

[0014] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the respiratory anomaly detection method based on gas monitoring as described in the first aspect of the present invention.

[0015] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the gas monitoring-based respiratory abnormality detection method as described in the first aspect of the present invention.

[0016] The beneficial effects of this invention are as follows: By combining differential analysis of multidimensional gas components and deep learning models with a personalized breathing pattern library, this invention achieves more accurate breathing pattern recognition and anomaly detection, thereby providing a more intelligent health management solution, which has significant advantages, especially in dealing with complex individual differences and long-term health trends. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. 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.

[0018] Figure 1 This is a flowchart of the respiratory anomaly detection method based on gas monitoring in Example 1.

[0019] Figure 2 This is a structural diagram of the respiratory anomaly detection system based on gas monitoring in Example 1. Detailed Implementation

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0021] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0022] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0023] Example 1, referring to Figure 1 and Figure 2 This is the first embodiment of the present invention, which provides a method for detecting abnormal breathing based on gas monitoring, including the following steps: S1. Collect and preprocess the gas signals from the user's mouth and nose during breathing, and separate the gas components through the microfluidic gas separation module; Specifically, collecting and preprocessing the gas signals from the user's mouth and nose during breathing involves the user wearing a dedicated flexible dual-channel wearable gas collection device. This device contains two independent gas sampling ports, corresponding to the collection of gas from the nasal cavity and oral cavity, respectively, and integrates a microfluidic gas separation chip and a high-precision gas sensor. The high-precision gas sensor includes a carbon dioxide sensor, an oxygen sensor, and a volatile organic compound sensor; The volatile organic compounds (VOCs) refer to organic gas molecules present in trace amounts in respiratory gases, such as ethanol, acetone, and isoprene, which may originate from metabolism, oral flora, or the external environment. VOC sensors typically employ metal oxide semiconductor or photoelectro-ionization technology to identify trace components in gas mixtures. When the device is started, it first automatically pairs and establishes a connection with the host computer system or mobile application via Bluetooth module. After the device is started, the two independent channels of nasal cavity and oral cavity will start to sample gas separately, acquiring 10 sets of raw data containing timestamps and gas concentrations per second. Each set of data contains real-time concentration readings of carbon dioxide, oxygen and volatile organic compounds in nasal cavity and oral cavity. High-pass filters are used to remove high-frequency noise from the gas signal. Fast Fourier transform is then used to perform frequency domain analysis on the noise-removed data, converting the time-domain gas signal into frequency-domain data to obtain the dominant frequency of respiration, spectral energy distribution, and amplitude characteristics of gas concentration changes.

[0024] By independently collecting nasal and oral air samples using a flexible dual-channel wearable device, and combining frequency domain analysis and differential analysis techniques, multi-dimensional and multi-level respiratory health data is provided. This not only offers a new method for early detection of respiratory abnormalities but also provides more reliable data support for personalized health management.

[0025] Furthermore, a fast Fourier transform is performed on the respiratory signal sequence to obtain the corresponding spectral distribution data, where the horizontal axis of the spectrum is frequency and the vertical axis is the corresponding energy / amplitude. The amplitude of the spectral distribution data is normalized to eliminate the influence of individual differences, and high-frequency noise components are weakened by smoothing filtering methods, making the spectral peak characteristics more significant. Local maxima are detected on the smoothed spectral curve, and the frequencies corresponding to each maxima are recorded to form a candidate peak set. Based on the amplitude, frequency position and relative energy ratio of the candidate peaks, false peaks are eliminated, and the peak with the largest amplitude within the normal respiratory rate range is determined as the dominant expiratory frequency peak. The determined peak frequency value of the expiratory main frequency and its corresponding amplitude parameters are output as respiratory characteristic indicators for subsequent analysis. The normal respiratory rate range is approximately 0.2 to 0.5 Hz (12 to 30 breaths per minute). When an exhalation peak is detected, a micro gas guiding device (such as a piezoelectric micropump or electronic valve) will be automatically activated to guide the nasal and oral gas in the current channel into the microfluidic separation chip respectively. At this time, the data of the two channels are synchronized and both have pre-processed tags for subsequent traceability and channel differentiation. After the gas enters the microfluidic chip, it first flows into a micro-region called the "laminar inlet chamber". This region strictly controls the gas inlet speed and channel size to keep the gas at a low Reynolds number (e.g., Re < 1) in the flow state. Re refers to the Reynolds number, which means that it is in a laminar rather than turbulent state. This ensures that the behavior of various gas components in the channel conforms to the controllable diffusion model. In the laminar state, the gas continues to enter the chip's "diffusion separation region" along the main channel. In the diffusion separation region, the channel cross-section is typically only tens of micrometers, and the mutual diffusion between gas molecules becomes the dominant mechanism. At this time, different gas components exhibit different diffusion rates due to their different molecular sizes. For example, oxygen has a small molecular weight and diffuses quickly, carbon dioxide has a medium molecular weight and diffuses at a moderate rate, while volatile organic compounds (VOCs) usually have a more complex molecular structure and diffuse more slowly. Due to the difference in diffusion rates, these gases gradually form relatively separated distribution regions in the transverse space of the channel during the flow process. The chip has multiple predefined outlets at the end of this separation channel. The location of each outlet corresponds to a high concentration area of ​​a gas in the lateral channel. These outlets guide specific gas components into their respective detection chambers through tiny lateral drainage structures. For example, oxygen diffuses faster and is closer to the channel edge, so its dedicated outlet is located on the outside, while VOCs diffuse more slowly and are concentrated in the center of the fluid, so their outlet is located in the middle of the main channel. This spatial separation method does not rely on chemical reactions or complex controls, but can be achieved solely through physical diffusion mechanisms, ensuring that the separation process is fast, low-energy, and continuously controllable. Each detection chamber is equipped with a highly sensitive MEMS gas sensor or electrochemical detection element. These sensors start real-time measurement immediately after the gas enters and record the concentration change of their respective target gas at a frequency of 10Hz or higher. The measurement data are all timestamped and retain the source label information of the nasal cavity and oral cavity for subsequent differential analysis. Finally, the detection results are fused with the original respiratory signal that was initially processed by high-pass filtering and Fourier transform to form a complete set of data features. This set includes the physical concentration values ​​of different gas components, time information, source channel identifiers, and their correspondence with the original signal.

[0026] Through the diffusion separation region within the microfluidic chip, gas components are effectively separated based on differences in molecular size and diffusion rate. Gas components such as oxygen, carbon dioxide, and volatile organic compounds (VOCs) can naturally separate spatially within the same channel due to their different diffusion rates. This physical diffusion-based separation method does not rely on complex chemical reactions or energy consumption, offering advantages such as speed, low energy consumption, and continuous controllability. The laminar flow design ensures the stability of the gas flow, allowing gas molecules to naturally distribute according to their respective diffusion rates. At low Reynolds numbers, gas molecules do not experience drastic disturbances, thus guaranteeing precise separation of gas components. This design avoids complex flow control and energy consumption, improving the reliability and performance stability of the device.

[0027] S2. Perform differential analysis on the separated gas components, calculate the fluctuation of gas components in the nasal cavity and oral cavity, and identify breathing patterns through 1D SNN. At the same time, establish a personalized breathing pattern library by combining the user's long-term health data. Specifically, differential analysis is performed on the separated gas components to calculate the fluctuations in the gas components of the nasal cavity and oral cavity. This involves performing differential analysis on the gas component signals from the nasal cavity and oral cavity, calculating the component differences between the two at the same time point, and obtaining the fluctuations in the gas components. The differential calculation formula is as follows: In the formula, It is the fluctuation of the gas composition in the nasal cavity and oral cavity. and These represent the gas component concentrations in the nasal cavity and oral cavity at time T, respectively. The root mean square value is used to calculate the amplitude of the differentially analyzed gas composition fluctuation signal to quantify the degree of fluctuation. In the formula, RMS is the root mean square value of the differential signal, representing the average amplitude of gas composition fluctuations, and N is the number of samples in the differential signal calculation.

[0028] Differential analysis of the gas composition in the nasal cavity and oral cavity can quantify the fluctuations between them, revealing changes in different gas components during respiration. This process can effectively extract features of breathing patterns, helping to identify different breathing patterns.

[0029] Furthermore, breathing pattern recognition is performed using 1D SNN, and a personalized breathing pattern library is established by combining users' long-term health data. This involves constructing a pattern recognition model based on the 1D SNN model, including building a time series matrix based on the obtained differential fluctuation data and converting the matrix into one-dimensional sequence segments to form candidate sample pairs. Two 1DCNN sub-networks with shared weights are used as the backbone of the Siamese structure. Each sub-network takes a gas fluctuation sequence segment as input and outputs its embedded feature vector. A distance calculation layer (such as Euclidean distance or Manhattan distance) is added on top to measure the similarity between the two sequences. Extract breathing pattern fragments (such as slow deep breathing, mouth breathing, high-frequency breathing, etc.) from the user's historical breathing monitoring data (from health records, sleep records, breathing during exercise, breathing patterns during illness, etc.), pair them with the current differential fluctuation data, and mark which belong to similar (same type of breathing) and which belong to dissimilar (different type of breathing). Supervised training is performed using similar-dissimilar paired samples. The sub-network learns the embedding spatial distribution between different breathing patterns. Data augmentation strategies (such as adding noise, scaling, and temporal misalignment) are introduced during training to enhance the generalization of the model, resulting in a well-trained pattern recognition model. Real-time differential gas fluctuation sequence segments are captured by a sliding window and fed into a pattern recognition model one by one for processing to obtain the corresponding breathing patterns, including but not limited to normal breathing, slow deep breathing, rapid shallow breathing, mouth breathing, and intermittent breathing. Collect users' historical breathing-related data, including sleep breathing monitoring (such as nighttime snoring and sleep apnea), daily activity breathing (exercise, rest, anxiety, etc.), medical and health records (such as lung disease, allergy, and asthma attack records), and health device data (such as smart bracelets and respiratory rate sensors). After standardization, perform pattern recognition and store the data in a personalized database. The data storage format is [breathing pattern type, embedding vector, timestamp, associated health status (such as after exercise, during a cold)]. For each new segment, a similarity calculation is performed with the existing embedding vectors in the database. If the similarity is higher than a set threshold A, it is determined to be a known breathing pattern; otherwise, it is determined to be a "new pattern" and enters the labeling queue. This includes visually displaying the new pattern to the user or doctor for confirmation, and using contextual information (such as the user being exercised / having a cold) to assist in automatic labeling. If the confirmation is valid, it is added to the pattern library as a new category expansion.

[0030] Using 1D SNNs for breathing pattern recognition can effectively capture the differences between various breathing patterns (such as slow deep breathing, rapid shallow breathing, and mouth breathing). Compared to traditional pattern recognition methods, SNNs, through sub-networks with shared weights, can better handle time-series data and learn the embedding features of different breathing patterns, thereby improving the accuracy and flexibility of pattern recognition. Furthermore, by introducing data augmentation strategies, the generalization ability of 1D SNN models is significantly improved, enabling them to better handle various individual differences and breathing abnormalities. Establishing a personalized breathing pattern library allows for the tailoring of breathing pattern recognition models based on each user's long-term health data. By combining historical data (such as sleep apnea monitoring and breathing patterns during exercise) with real-time gas signals and continuously updating the pattern library, the accuracy and personalization of breathing pattern recognition can be improved. This library can adjust the judgment criteria for breathing patterns according to the user's specific situation (such as after exercise or during illness), thereby achieving more accurate abnormality detection and health management.

[0031] S3. Detect anomalies in real-time gas signals and compare them with a personalized breathing pattern library. If an anomaly is detected, issue an early warning notification immediately, activate the breathing assistance mode, and generate personalized health recommendations. Specifically, anomaly detection of real-time gas signals and comparison with a personalized breathing pattern library include: Historical gas signals are collected to generate historical time series data. Each data point includes multiple gas components (such as oxygen concentration, carbon dioxide concentration, etc.). The time series data is divided into data segments of length 200 using the sliding window method, i.e., each segment contains 200 data points and corresponds to 2 seconds in length (assuming a sampling frequency of 100Hz). Each time window represents an input sample. All sample segments are standardized (i.e., normalized) to ensure that the feature values ​​of each time window are on a uniform scale, avoiding certain features from dominating training. All standardized time segments are stored as a training dataset and processed in batches for input into the neural network for training. For each input time segment, a maximum overlap discrete wavelet transform is performed. The Coiflet 5 wavelet function is selected for signal decomposition. The wavelet decomposition level J=3, that is, the signal is decomposed at 3 levels, resulting in the following wavelet coefficients: W1(t): Level 1 detail coefficient, representing high-frequency components, capable of capturing rapid fluctuations (such as short-term changes in breathing). W2(t): Level 2 detail coefficients, capturing medium-frequency fluctuations at lower frequencies; W3(t): The third level detail coefficient, which captures low-frequency fluctuations and reflects slower changes; V3(t): The third-level approximation coefficient, representing the low-frequency components of the signal, which can capture the overall breathing rhythm (e.g., slow, deep breathing). Obtain the wavelet coefficients of each layer : The wavelet coefficients of each layer are concatenated into an enhanced feature vector X. t ; An encoder network is constructed using LSTM layers to enhance the input feature vector X. t Extract temporal features and generate a low-dimensional latent space representation Z. t This includes the following steps: A three-layer bidirectional LSTN (256 units per layer, ReLU activation function) is used to process the enhanced multi-scale feature X. t : The first LSTM layer: processes the forward and reverse dependencies of time series, with X as the input and the hidden state of the first layer as the output; The second LSTM layer: Based on the output of the first layer, further extract temporal features and perform forward and reverse processing in the same way; The third LSTM layer further deepens feature extraction, mapping the captured temporal information into the final low-dimensional embedding representation; The number of LSTM units in each layer is set to 256, and ReLU is used as the activation function to increase the nonlinear expressive power of the network. Bidirectional LSTM networks can simultaneously consider past and future information of the input sequence, which is especially important for time series data because certain features may affect the current state at future moments. To learn temporal dependencies, each layer of the bidirectional LSTM processes both forward and backward time series information, ensuring that it can capture the dependencies between consecutive time points. The last LSTM layer of the encoder outputs the hidden state h. t This hidden state h t Z, as a low-dimensional latent space representation t : In the formula, It is a time-series input enhanced by MODWT. It is the encoder part in the LSTM Autoencoder model, which consists of a series of LSTM layers; The LSTM decoder maps the low-dimensional latent space representation generated by the encoder back to the original signal space, reconstructing an approximation of the input signal. Specifically, this includes: The input to the decoder is the low-dimensional latent representation Z of the encoder output. t This part is a condensed version of the temporal information extracted by the encoder, which contains a high-level feature representation of the original input data; The decoder consists of three LSTM layers and is symmetrical to the encoder structure. First-layer LSTM: Represents the latent space Z t Decode it into a new temporal feature vector; The second LSTM layer further maps the output of the first layer to higher-level temporal features; The third LSTM layer: ultimately maps the high-level temporal features back to the original signal space, generating a reconstructed signal; Each layer of the decoder also uses 256 LSTM units with ReLU activation function to ensure that the network has sufficient expressive power. The decoder output is That is, the signal generated by the model through the decoding process serves as a reconstructed version of the original signal: In the formula, It is the time series reconstructed by the decoder. It is the decoder part of the LSTM Autoencoder; By optimizing the loss function Minimize reconstruction error: In the formula, m is the number of time steps in a single sample of the time series. and These are the input values ​​of the original signal and the reconstructed signal at the j-th time point, respectively; To enable the LSTM-Autoencoder model to reconstruct results that are as close as possible to the input signal, the model learns to extract effective temporal features from the input signal and map them back to the original signal space by continuously adjusting the network parameters. The Adam optimizer is used to update the network weights, with a learning rate of 0.0001, a batch size of 32, and 100 training epochs. To prevent overfitting, an early stopping strategy is adopted, automatically stopping training when the loss on the validation set no longer improves. During training, the loss values ​​on the training and validation sets are evaluated periodically to ensure that the model can converge on the training data and maintain good generalization ability on the validation set. By observing the changing trends of the training and validation losses, overfitting can be effectively prevented and the training strategy can be adjusted to obtain a well-trained LSTM-Autoencoder model. The real-time enhanced feature vectors are input into the trained LSTM-Autoencoder model to obtain the reconstructed signal. The original input signal and the reconstructed signal output by the model are compared, and the reconstruction error is calculated using Euclidean distance (L2 norm). : Using the reconstruction error on the training set, calculate the mean and standard deviation of the error: In the formula, It is the mean of the reconstruction error of the training set. is the standard deviation of the reconstruction error of the training set, and n is the number of training samples (i.e., the total number of sample segments). It is the reconstruction error of the i-th training sample; Based on the statistical characteristics of the training set, a dynamic anomaly detection threshold θ is set: In the formula, k is the anomaly tolerance coefficient, which is set through cross-validation; If the reconstruction error at the current moment is greater than the threshold θ, the sample is judged as abnormal. Based on the reconstruction error judgment, the embedding features Z of the encoder output are used by One-Class SVM. t Boundary strengthening is performed to further confirm anomalous samples, specifically including: Use the encoder output corresponding to all normal samples in the training set. Construct a feature set, train a One-Class SVM using the RBF kernel function, construct the boundary hyperplane for normal samples, and obtain the boundary function: In the formula, The score is determined using OC-SVM. These are the Lagrange multipliers corresponding to the SVM support vectors, obtained through internal model optimization during training. These are support vector feature points, which are the embedding features of normal samples retained in the training set. It is the bias term in the SVM model, determined during model training, and returned internally by sklearn. It is a kernel function (such as the RBF kernel) that measures the non-linear similarity between input features and support vectors. It is the kernel width parameter of the RBF kernel function, which controls the tightness of the support vector boundary and is automatically tuned using grid search cross-validation; The latent space representation Z corresponding to the current real-time signal is... t Inputting the data into a pre-loaded One-Class SVM model yields the scalar value f(Z). t ), indicating whether the sample is within the normal boundary learned from the training samples, if f(Z) t If )≥0, it means the sample is inside the normal boundary; otherwise, it means the sample is outside the boundary and is judged as abnormal. After completing the boundary determination, the latent spatial representation vector Z corresponding to the time segment is obtained. t That is, the low-dimensional embedding feature representation output from the encoder part of the LSTM-Autoencoder, which will be Z t As input, a personalized breathing pattern library is queried for comparison. This library consists of breathing data from the user's historical health status. After being decomposed by MODWT and extracted by the encoder, it is also transformed into a series of embedding vectors with scene labels (such as "sleep", "exercise", "fever state" etc.). During the comparison process, the similarity between the current vector and each reference vector in the library is calculated one by one using Euclidean distance or cosine similarity. If the similarity between the current embedding vector and one or more human body patterns in the library exceeds a preset threshold, the current signal is determined to be a "known normal behavior pattern"; otherwise, it is marked as an "unknown behavior pattern". The mechanism integrates three judgment results: whether the reconstruction error exceeds the threshold, whether the One-Class SVM judgment goes out of bounds, and whether it matches the personalized pattern library. The combination of these three indicators constitutes the basis for the joint anomaly judgment mechanism. Table 1: Example Table of Joint Anomaly Detection Mechanism The samples are divided into different levels: mild abnormality, moderate abnormality, severe abnormality, and new patterns to be confirmed, and can automatically identify newly emerging behavioral patterns.

[0032] MODWT technology effectively removes noise from signals through multi-level wavelet decomposition, preserving the effective characteristics of gas composition fluctuations and enhancing the quality of the input signal, providing a clean data foundation for subsequent processing. Compared with traditional signal processing methods, MODWT can more accurately extract multi-frequency features from complex signals, and is particularly suitable for the analysis of non-stationary and irregular gas signals. The LSTM-Autoencoder model can extract key temporal features from the enhanced signal, making it particularly suitable for processing long-term series data. The long short-term memory capability of LSTM effectively captures the long-term dependence of respiratory signals, while the autoencoder further reduces the impact of noise and improves the accuracy of respiratory pattern recognition through low-dimensional latent space mapping and signal reconstruction. One-Class SVM further enhances the ability to detect anomalies. By establishing the boundary of normal patterns, One-Class SVM can efficiently distinguish abnormal signals, especially for novel or unknown abnormal patterns, exhibiting strong adaptability and reliability.

[0033] The combination of these three technologies results in a multi-layered, comprehensive anomaly detection mechanism. MODWT first filters out unnecessary noise, ensuring the quality of the input signal; LSTM-Autoencoder performs efficient feature extraction and signal reconstruction in the context of time-series data, capturing key temporal change information; One-Class SVM further strengthens the boundary judgment of anomalous samples, ensuring more accurate detection of anomalous signals. Based on this, through triple verification of reconstruction error, SVM boundary detection, and pattern library comparison, the system can comprehensively judge signals from multiple perspectives, ensuring higher detection accuracy and reliability. This multi-layered anomaly detection mechanism is unparalleled by traditional technologies. Traditional anomaly detection often relies on a single model or method, is easily affected by noise interference, or is not sensitive enough to complex, long-term accumulated anomaly patterns. This invention, by integrating MODWT, LSTM-Autoencoder, and One-Class SVM, overcomes the limitations of single technologies, thereby effectively improving the system's anomaly detection capability, accuracy, and adaptability.

[0034] Furthermore, activating the respiratory support mode and generating personalized health recommendations includes: When the joint anomaly determination mechanism outputs a severe anomaly or a new mode to be confirmed and the anomaly type in the personalized breathing mode library is a severe respiratory anomaly (e.g., apnea, rapid breathing, abnormal respiratory rate, etc.), it is determined that the respiratory status is critical and the "respiratory support mode" is immediately triggered. The breathing assistance mode includes guiding users to relax and relieve tension through visual (such as soothing graphics or animations on the interface) or auditory (such as soothing music, deep breathing guidance, etc.), providing short-term breathing training or meditation guidance to help users slow down their breathing and restore their rhythm, controlling the smart device and mask to continuously adjust airflow and pressure, while continuing to help users adjust their state through interface and voice prompts, the device and system monitor the user's reaction in real time, and adjust the assistance intensity and mode through feedback to ensure that the user restores a normal breathing rhythm; Collect and integrate user health data and environmental information, including real-time respiratory data (e.g., respiratory rate, respiratory depth, apnea time, breathing pattern type, etc.). Physiological data (e.g., heart rate, blood oxygen saturation, body temperature, etc.); Environmental data (e.g., air quality, humidity, temperature, noise, etc.); User's emotional state data (e.g., anxiety level, relaxation level, mood changes, etc.); The collected data is preprocessed, including standardization: feature data from different sources (such as respiratory data, environmental data, emotional state, etc.) are standardized to ensure that they are within the same scale and are easy to input into the model; Feature fusion: This involves fusing various types of data (such as respiratory data, health data, and emotion data) into a comprehensive feature vector, which serves as the input for subsequent algorithms. Continuous health data (such as heart rate, respiratory rate, blood oxygen saturation, etc.) is divided into small segments by time windows. Each segment represents a time period (such as every 1 minute or 30 seconds). For each time window, statistical characteristics such as mean, standard deviation, and trend are calculated. For example, the mean, standard deviation, and trend of respiratory rate per minute are calculated. These statistical characteristics (such as mean, standard deviation, and trend) are combined into a feature vector, which is used as the representation of the time window. The feature vector of each time window reflects the user's health status during that period. Multiple segment-level feature vectors are concatenated or fused to form a period-level health status representation. For example, multiple 1-minute health status vectors from a day are concatenated to form a 1-hour or 1-day health status representation. By processing the spliced ​​segment features through LSTM or GRU encoding networks, a periodic health status vector is generated. This step helps to capture the health trends and fluctuations throughout the entire period. The weight of each segment or periodic feature is calculated through an attention mechanism. For example, if certain abnormal moments (such as sleep apnea) are critical to health status, then the weight of that moment will be higher. Based on the calculated weights, health features are weighted to ensure that the system focuses on abnormal situations (such as sudden heart rate fluctuations) while ignoring routine health status, resulting in a weighted health status vector, which further strengthens the key moments and fluctuation information in health status analysis. A preliminary health management plan is generated based on the weighted health status vector; A hybrid optimization algorithm, WCSO, is employed to iteratively optimize the health management program by combining the global search capability of the water cycle algorithm with the local refinement capability of the competitive swarm optimization algorithm. Specifically, this includes: Initialize the population. Each individual in the population represents a health management plan. Each health management plan includes multiple health recommendations (such as diet, exercise, rest, sleep, etc.). For example, a health management program might include: Recommend how many minutes of aerobic exercise users should engage in; It is recommended to adjust sleep duration; Suggestions for increasing or decreasing nutrient intake, etc. Each health management plan includes multiple dimensions, each of which can be generated by weighting different health data features (such as respiratory rate, blood oxygen saturation, heart rate, emotional state, etc.). WCSO initializes the population based on these features to form different health management plans. The algorithm extracts potential optimal solutions from the current health management plan through evaporation. Specifically, it randomly selects some potential better solutions from the initial health management plans (the selection probability of each initial plan is related to its effect on health management (e.g., improving breathing, promoting exercise, etc.). The algorithm prioritizes health management plans with greater potential for improvement in both short-term and long-term health). The extracted solutions are then condensed to screen out solutions with high potential. These solutions with higher potential are considered more suitable for the user's health management needs. During precipitation, multiple optimal solutions are merged to further search for the optimal health management plan, avoiding local optima and ensuring that the global optimum is found, thus laying the foundation for subsequent local optimization. The population of health management programs optimized by WCA is divided into multiple subpopulations using CSO. Each subpopulation contains H candidate health management programs, and each subpopulation represents a different optimization goal and strategy. Competition takes place within each subgroup to select the best-performing individual. An individual's performance can be evaluated using the following criteria: Health benefits: such as whether the amount of exercise meets the user's health goals (e.g., improving cardiopulmonary function), whether sleep quality is improved, and whether the diet is reasonable; Optimization goal: To evaluate the effectiveness of the plan using certain indicators (such as maximizing the overall improvement in health status); Selection process: For each subgroup, the best-performing individuals (health management plan) are selected. These individuals will become the "leaders" of that subgroup, meaning their characteristics will provide direction for other individuals to learn from and adjust. During the competition, each individual (health management plan) in the subgroup adjusts its strategy based on the performance of the best individual. For example, if a plan involves less exercise, while the winning plan in the competition performs well in terms of exercise recommendations, other individuals may adjust their exercise time to make adaptive optimizations. The optimal health management plan output by WCA is locally optimized through CSO. After local optimization, the health management plan of all individuals will be updated according to the new optimization results. Each plan will be adjusted and evolved according to the feedback of the local optimization steps. The updated individuals will return to the population and become new population members. At this time, the individuals in the population will have more diversity and will be closer to the global optimum. According to the set stopping conditions (such as the maximum number of iterations or reaching the satisfaction standard), the system will continue to optimize until the health management plan achieves the expected effect or meets the user's needs. The local optimization content is as follows: Exercise optimization: For example, if a health management plan does not provide enough exercise time, the system will adjust the type or intensity of exercise without significantly increasing the time burden, in order to improve the user's overall health level. Sleep optimization: If sleep time is too long or too short, the system will optimize sleep time and quality based on the user's health data and goals, such as extending deep sleep time or suggesting appropriate adjustments to nighttime sleep time; Diet optimization: By adjusting the proportions of dietary components (such as reducing sugar intake and increasing vegetable intake), users' dietary structure is made more in line with health goals; By making small adjustments and making local optimizations, the effectiveness can be improved by fine-tuning health recommendations. For example, adding 5 minutes of aerobic exercise to an exercise recommendation, or adjusting a sleep recommendation to increase deep sleep time by 10 minutes. If there are multiple goals in the health management plan (such as increasing exercise and adjusting diet), the CSO will perform multi-goal optimization based on the priority and weight of the goals to ensure that the optimized plan takes into account the needs of all aspects. Individuals in a subgroup not only learn by competing with the best individual, but also share optimization information with other individuals and learn from each other's optimization experience. Through collective cooperation, multiple individuals can share the optimal solution in a local area, thereby improving the overall optimization effect. Through localized optimization, health management plans will become more personalized, enabling precise adjustments to meet specific user needs and improve health management effectiveness. Based on the optimization schemes output by WCA and CSO, and combined with the user's health status vector, different weights are assigned to each health suggestion. Personalized health management suggestions are generated based on the weighted schemes, including short-term suggestions (such as "increase walking time" and "do deep breathing exercises") and long-term suggestions (such as "improve sleep quality" and "increase aerobic exercise"). The generated personalized health suggestions are pushed to the user, and the health management plan is dynamically adjusted based on user feedback.

[0035] Introducing an attention mechanism into LSTM or GRU models allows the model to automatically focus on time windows or features that significantly impact a user's health. For example, sudden changes in respiratory abnormalities are often confined to specific time periods; the attention mechanism helps the model assign higher weights to these critical moments, significantly improving its response to sudden anomalies. This mechanism enables the system to weight data based on key components of the time series, enhancing the accuracy of identifying complex breathing patterns. WCSO combines the global search capability of WCA (Water Cycle Algorithm) with the local optimization capability of CSO (Competitive Group Optimization), allowing the system to search a broad solution space while finely adjusting the weights and strategies of each health suggestion when generating personalized health management plans. WCA extensively explores the solution space during the initial plan generation process to ensure the discovery of potential optimal solutions; while CSO refines the plans through competition and local optimization, ensuring that the generated health management plans are more personalized and accurate. Compared to traditional optimization methods, WCSO's hybrid optimization provides health management plans for each user more efficiently and accurately, especially demonstrating stronger advantages when handling multi-objective, multi-dimensional optimization tasks.

[0036] This embodiment also provides a respiratory anomaly detection system based on gas monitoring, including: The gas signal acquisition module is used to acquire the user's oral and nasal gas signals, preprocess them, remove noise, and perform frequency domain analysis. A microfluidic gas separation module is used to separate the gas components in the oral cavity and nasal cavity using microfluidic technology; The differential analysis module is used to calculate the fluctuations in the gas composition of the nasal cavity and oral cavity, and to quantify the differences in gas composition. The breathing pattern recognition module is used to recognize breathing patterns using 1D SNN and to build a personalized breathing pattern library by combining user health data. The health advice generation module is used to detect anomalies in real-time gas signals and compare them with a personalized breathing pattern library. Based on the anomaly detection results, it activates the breathing assistance mode and generates personalized health advice.

[0037] This embodiment also provides a computer device applicable to the respiratory anomaly detection method based on gas monitoring, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the respiratory anomaly detection method based on gas monitoring as proposed in the above embodiment.

[0038] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0039] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the respiratory anomaly detection method based on gas monitoring as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0040] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method of breath abnormality detection based on gas monitoring, the method comprising: The application relates to a real-time respiratory state monitoring and health management system. ​ The application comprises the following steps: Collecting and preprocessing the oral-nasal gas signals during the user's breathing process, separating the gas components through a microfluidic gas separation module; Performing differential analysis on the separated gas components, calculating the fluctuations of the nasal and oral gas components, and performing respiratory pattern recognition through a 1D SNN, and establishing a personalized respiratory pattern library by combining the user's long-term health data; 2. The gas monitoring based respiratory abnormality detection method as claimed in claim 1, wherein: Detecting abnormalities in real-time gas signals and comparing them with the personalized respiratory pattern library, issuing an early warning notification if an abnormality is detected, starting a respiratory assistance mode and generating personalized health recommendations. The respiratory pattern recognition through a 1D SNN, while combining the user's long-term health data to establish a personalized respiratory pattern library, is based on a 1D SNN model to construct a pattern recognition model; The real-time differential gas fluctuation sequence fragments are intercepted through a sliding window and are sent one by one into the pattern recognition model for processing to obtain the corresponding respiratory pattern; 3. The gas monitoring based breath abnormality detection method of claim 2, wherein: Collecting the user's historical respiratory-related data, also standardizing and performing pattern recognition, and storing them into the personalized database, and calculating the similarity with the existing embedded vectors in the database, if the similarity is higher than the set threshold A, it is determined as a known respiratory pattern, otherwise it is determined as a new pattern and enters the marking queue. Collect historical gas signals, generate historical time series data, cut time series data into data segments using a sliding window method, each time window represents an input sample, standardize all sample segments, store all standardized time segments as a training data set, perform maximum overlap discrete wavelet transform on each input time segment, select Coiflet 5 wavelet function for signal decomposition, set wavelet decomposition level, and obtain wavelet coefficients at each level Splice the wavelet coefficients at each level into an enhanced feature vector X t ; An encoder network is constructed using LSTM layers to extract temporal features from the input augmented feature vector X t , generating a low-dimensional latent space representation Z t An LSTM decoder is used to map the low-dimensional latent space representation generated by the encoder back to the original signal space, reconstructing an approximation of the input signal, resulting in ; By optimizing the loss function Minimizing reconstruction error, making the LSTM-Autoencoder model able to reconstruct the results as close as possible to the input signal, by constantly adjusting the parameters of the network, the model will learn to extract effective time series features from the input signal and map them back to the original signal space, using the Adam optimizer to update the network weights, using the early stopping strategy, when the loss of the validation set no longer improves, automatically stop training, get the trained LSTM-Autoencoder model; The real-time enhanced feature vector is input into the trained LSTM-Autoencoder model to obtain a reconstructed signal, the original input signal and the reconstructed signal output by the model are compared, and the reconstruction error is calculated using the Euclidean distance , a dynamic anomaly determination threshold θ is set, if the reconstruction error at the current time is greater than the threshold θ, the sample is determined as an anomaly, on the basis of the reconstruction error judgment, the embedding feature Z output by the encoder is used as input to query the personalized respiratory mode library for comparison t , boundary reinforcement is performed to further confirm the abnormal samples, and Z t is used as input to query the personalized respiratory mode library for comparison, in the comparison process, the similarity between the current vector and each reference vector in the library is calculated one by one through the Euclidean distance, if it is detected that the similarity between the current embedding vector and the mode in the library exceeds a preset threshold, it is determined that the current signal is a known normal behavior mode, otherwise it is marked as an unknown behavior mode; The abnormality detection of real-time gas signals and the comparison with the personalized respiratory pattern library include:

4. The gas monitoring based breath abnormality detection method of claim 3, wherein: Fusing the three judgment results, including whether the reconstruction error exceeds the threshold, whether the One-Class SVM determines the boundary, and whether it matches the personalized pattern library, the combination of the three indicators constitutes the basis of the joint abnormality judgment mechanism and divides the samples into different levels, including mild abnormality, moderate abnormality, severe abnormality and new pattern to be confirmed. The start of the respiratory assistance mode and the generation of the personalized health recommendations include: When the joint abnormality judgment mechanism outputs severe abnormality or new pattern to be confirmed and the abnormality type in the personalized respiratory pattern library is a serious respiratory abnormality, it is determined that the respiratory state is critical, and the "respiratory assistance mode" is triggered immediately; Collecting and integrating the user's health data and environmental information and preprocessing, dividing the continuous health data into small fragments according to the time window, calculating the statistical features in each time window, combining the statistical features into a feature vector, splicing the feature vectors at the fragment level to form the health state representation at the period level and generate the period-level health state vector, calculating the weight of each fragment or period feature through the attention mechanism, obtaining the weighted health state vector and generating the preliminary health management scheme; Using the hybrid optimization algorithm WCSO, combining the global search ability of the water cycle algorithm and the local fine-tuning ability of the competitive swarm optimization algorithm, iteratively optimizing the health management scheme, according to the optimization scheme output by WCA and CSO, combining the user's health state vector, assigning different weights to each health recommendation, generating personalized health management recommendations according to the weighted scheme, including short-term recommendations and long-term recommendations, pushing the generated personalized health recommendations to the user, and dynamically adjusting the health management scheme according to the user feedback.

5. The gas monitoring based breath abnormality detection method of claim 4, wherein: The differential analysis of the separated gas components calculates the fluctuation of the nasal and oral gas components.

6. The gas monitoring based breath abnormality detection method of claim 5, wherein: The collection of the oral and nasal gas signals during the user's breathing and the preprocessing thereof collect the raw data of the nasal and oral gas concentrations, including the real-time concentration readings of carbon dioxide, oxygen, and volatile organic compounds in the nasal and oral cavities. The high-frequency noise is removed from the gas signals using a high-pass filter, and the frequency domain analysis is performed on the noise-removed data through fast Fourier transform.

7. The gas monitoring based breath abnormality detection method of claim 6, wherein: The gas component separation through the microfluidic gas separation module involves positioning the exhalation main frequency peak based on the fast Fourier transform result, and introducing the nasal and oral gases in the current channel into the microfluidic separation chip. The gas is subjected to preliminary flow adjustment to ensure that the fluid is in a stable laminar flow state, and then the gas flows into the separation channel. Under the guidance of multiple transverse micro-branch channels, the component diffusion separation process is performed, thereby obtaining the separated gas components.

8. A respiratory abnormality detection system based on gas monitoring, based on the respiratory abnormality detection method based on gas monitoring according to any one of claims 1 to 7, characterized in that: The method comprises the following steps: a gas signal acquisition module for acquiring the oral and nasal gas signals of the user and preprocessing the same, removing noise, and performing frequency domain analysis; a microfluidic gas separation module for separating the oral and nasal gas components through microfluidic technology; a differential analysis module for calculating the fluctuation of the nasal and oral gas components and quantifying the gas component difference; a respiratory pattern recognition module for recognizing the respiratory pattern using 1D SNN and establishing a personalized respiratory pattern library in combination with the user's health data; a health suggestion generation module for detecting the abnormality of the real-time gas signal, comparing it with the personalized respiratory pattern library, starting the respiratory assistance mode according to the abnormality detection result, and generating personalized health suggestions. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is characterized in that: The processor executes the computer program to realize the steps of the respiratory abnormality detection method based on gas monitoring according to any one of claims 1-7.

10. A computer readable storage medium having stored thereon a computer program, characterized in that: The computer program is executed by the processor to realize the steps of the respiratory abnormality detection method based on gas monitoring according to any one of claims 1-7.

Images