Remote monitoring method of portable intelligent breathing mask

Abstract

The invention relates to the technical field of medical health monitoring, in particular to a portable intelligent breathing mask and a remote monitoring method thereof. The method comprises the steps that airflow, humidity, blood oxygen and temperature data in the mask wearing process of a user are obtained, and time sequence alignment and feature fusion are carried out; inputting the fused physiological feature sequence into an embedded neural network model to recognize respiratory behaviors, and generating an abnormal probability value; carrying out compression coding based on the parameters, setting an uploading period according to a self-adaptive formula, and uploading the data to a cloud end; the cloud model further analyzes and generates health assessment parameters and a risk trend sequence; and dynamically adjusting an early warning threshold according to an evaluation result, and updating a local model and a sampling strategy in combination with user feedback. The portable breathing mask has the functions of fusing multi-modal data analysis, edge-cloud collaborative recognition, self-adaptive uploading scheduling, personalized early warning regulation and control and the like, and the intelligent level and practicability of the portable breathing mask in remote health monitoring are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical health monitoring, and in particular to a remote monitoring method for a portable intelligent breathing mask. Background Art

[0002] With the development of wearable health devices, respiratory monitoring technology is increasingly being used in scenarios such as chronic disease management, elderly care, and home rehabilitation. While traditional respiratory masks can assist breathing or filter gases, they typically only provide basic physical filtering functions and lack the ability to actively sense and intelligently respond to the user's real-time physiological state, making it difficult to dynamically identify abnormal respiratory states and provide personalized interventions.

[0003] Although some existing smart respiratory devices have introduced some sensor components for data collection, most of them have the following technical deficiencies: First, they rely only on a single physiological parameter (such as airflow or blood oxygen) for status monitoring, which is susceptible to noise interference and has limited judgment accuracy; second, they are unable to form a closed-loop structure from front-end data collection, edge recognition, remote analysis to system feedback, and lack the ability to respond to and control abnormal conditions in real time; third, the system cannot adaptively adjust the recognition model and sampling strategy based on individual differences and user feedback information, which can easily lead to excessive energy consumption or detection delays.

[0004] Therefore, there is an urgent need for an intelligent respiratory monitoring system that can integrate multimodal physiological signals such as airflow, humidity, blood oxygen saturation, and temperature, and combine remote health analysis results with user feedback information to dynamically identify abnormal respiratory states, realize personalized risk warnings, and adaptively control front-end acquisition strategies, so as to improve the practicality, accuracy, and response efficiency of the equipment. Summary of the Invention

[0005] The present invention provides a remote monitoring method for a portable intelligent respiratory mask to solve the problem of how to dynamically identify abnormal breathing status of users and realize personalized risk warning and adaptive regulation of front-end perception strategies based on multimodal physiological data such as airflow parameters, humidity parameters, blood oxygen saturation and temperature, combined with remote health analysis parameters and user interactive feedback information.

[0006] In order to solve the above technical problems, the present invention provides a portable intelligent breathing mask and a remote monitoring method thereof, comprising:

[0007] Acquiring multimodal raw physiological data generated by the user while wearing the mask, the raw physiological data including airflow parameters, humidity parameters, blood oxygen parameters, and temperature parameters, and performing timestamp alignment and feature fusion on the multimodal raw physiological data to obtain a physiological feature sequence in a unified format;

[0008] Inputting the physiological feature sequence into an embedded neural network model to perform respiratory behavior recognition to obtain a respiratory behavior label and an abnormal probability value parameter;

[0009] Based on the abnormal probability value parameter, the physiological feature sequence is compressed and coded to obtain compressed feature data, and a data upload cycle is set:

[0010] The compressed feature data and the respiratory behavior label are uploaded to the cloud platform via a low-power wireless communication module, and then input into a deep model built based on a bidirectional gated recurrent unit and an attention mechanism to generate a remote health assessment parameter and a sequence of the user's respiratory health risk trend;

[0011] generating alarm data and response level labels based on the remote health assessment parameters and the user's respiratory health risk trend sequence;

[0012] Based on the remote health assessment parameters and the interactive health feedback data set, the edge model and sensor sampling strategy are updated to achieve dynamic regulation of the front-end acquisition frequency and inference cycle.

[0013] Furthermore, the feature fusion includes channel normalization, feature extraction and feature fusion of the airflow parameters, humidity parameters, blood oxygen parameters and temperature parameters.

[0014] Furthermore, the embedded neural network model is a gated recurrent unit GRU and / or a long short-term memory network LSTM structure.

[0015] Furthermore, the compression coding process includes selecting at least one compression method of mean downsampling, principal component analysis PCA and / or lightweight lossless compression based on the abnormal probability value parameter.

[0016] Furthermore, the low-power communication module includes any one of a Bluetooth BLE module, an NB-IoT module and / or a LoRa module.

[0017] Furthermore, the deep model is a temporal modeling network constructed based on a bidirectional gated recurrent unit BiGRU and an attention mechanism.

[0018] Furthermore, the remote health assessment parameters include respiratory cycle expected value, fluctuation standard deviation, abnormality confidence and individual portrait correlation weight.

[0019] Furthermore, the alarm data is generated by weighted adjustment of static thresholds based on remote health assessment parameters and user respiratory health risk trend values, supporting three-level response level judgment.

[0020] Furthermore, the interactive health feedback data set is obtained based on user subjective state feedback data and auxiliary health record data.

[0021] Furthermore, the sensor sampling strategy includes sampling frequency, power consumption level and inference cycle configuration, and the dynamic regulation is jointly driven based on the remote health assessment parameters and the interactive health feedback dataset.

[0022] The present invention effectively enhances the synergy and temporal stability between data features and improves the accuracy of model recognition by uniformly aligning and weighting the mutual information of multimodal raw physiological data such as airflow, humidity, blood oxygen saturation and temperature; it uses an embedded neural network model to realize the preliminary recognition of respiratory behavior at the edge, with low latency and local reasoning capabilities, reducing dependence on network continuity; the present invention introduces an adaptive upload strategy based on abnormal probability value parameters for the first time. Through the above strategy, high-frequency upload is achieved under high-risk conditions, and the upload cycle is extended under low-risk conditions, taking into account both communication energy consumption and real-time performance; the cloud model constructed by combining BiGRU and attention mechanism performs personalized modeling analysis on uploaded data and user portraits to generate risk trend sequences; and through a dynamic multi-level early warning strategy, the local model parameters and sampling strategy are automatically updated in combination with user feedback data, realizing adaptive control of data acquisition frequency and model reasoning cycle. The above method effectively improves the data processing efficiency, risk identification accuracy and remote health management capabilities of the equipment, and is suitable for respiratory health monitoring in chronic disease populations, the elderly and special scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 A flowchart of a portable smart breathing mask and a remote monitoring method thereof provided in an embodiment of the present application;

[0024] Figure 2 This is a structural block diagram of the portable intelligent breathing mask and its remote monitoring system provided in the embodiments of the present application. DETAILED DESCRIPTION

[0025] Unless otherwise defined, all technical and scientific terms used herein have the same meanings as commonly understood by those skilled in the art to which this application belongs. The terms used in the specification of the application are for the purpose of describing specific embodiments only and are not intended to limit this application. The terms "including" and "having" and any variations thereof in the specification and claims of this application and the above-mentioned drawings are intended to cover non-exclusive inclusions. The terms "first", "second", etc. in the specification and claims of this application or the above-mentioned drawings are used to distinguish different objects, not to describe a specific order.

[0026] References herein to "embodiments" mean that a particular feature, structure, or characteristic described in connection with the embodiments may be included in at least one embodiment of the present application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it constitute an independent or alternative embodiment that is mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0027] Example 1: Reference Figure 1 , is a flow chart of a portable smart breathing mask and a remote monitoring method thereof provided by an embodiment of the present invention. The flow chart may include at least steps S100-S700:

[0028] S100, obtaining multimodal raw physiological data generated by a user while wearing a mask, the raw physiological data including airflow parameters, humidity parameters, blood oxygen parameters, and temperature parameters, performing time sequence alignment and feature fusion processing on the multimodal raw physiological data to obtain a physiological feature sequence in a unified format;

[0029] S200, inputting the physiological feature sequence into an embedded neural network model to perform preliminary identification of abnormal respiratory state, and obtaining a respiratory behavior label and an abnormal probability value parameter;

[0030] S300: Based on the abnormal probability value parameter, compress and encode the physiological feature sequence to obtain compressed feature data, set a data upload cycle according to the risk level of the compressed feature data, and upload the compressed data and the respiratory behavior tag to the cloud platform via a low-power wireless communication module;

[0031] S400: Obtain the compressed data and respiratory behavior labels uploaded to the cloud platform, combine them with historical health data and user profile information, input them into a deep model built based on a bidirectional gated recurrent unit and an attention mechanism, and generate a remote health assessment parameter and a user respiratory health risk trend sequence;

[0032] S500: Based on the remote health assessment parameters and the user's respiratory health risk trend sequence, dynamically adjust the preset threshold model, trigger the corresponding multi-level warning mechanism in combination with the latest abnormal judgment result, and generate alarm data and response level labels;

[0033] S600: Acquire the alarm data and risk trend sequence, generate remote visualization display content, and receive subjective status feedback data and auxiliary health record data uploaded by the user to form an interactive health feedback data set;

[0034] S700: Based on the remote health assessment parameters and the interactive health feedback data set, update the edge model parameters and sensor sampling strategy to complete the dynamic adjustment of the front-end data collection frequency and the edge model inference cycle;

[0035] Step S100 at least includes steps S110-S130:

[0036] S110: Obtain airflow parameters, humidity parameters, blood oxygen parameters, and temperature parameters generated when the user wears the mask, collect original time series data, and obtain multimodal original physiological data.

[0037] When the user wears the portable smart breathing mask, the airflow sensor, humidity sensor, blood oxygen sensor and temperature sensor installed inside the breathing mask are used to collect in-situ data. Specifically, the airflow sensor is used to obtain the change of gas flow rate per unit time in the breathing channel in real time, which is expressed as an airflow parameter sequence. The humidity sensor is used to obtain the water vapor concentration value in the user's exhaled gas, which is expressed as a humidity parameter sequence The blood oxygen sensor obtains a dynamic sequence indicating blood oxygen saturation, which is expressed as a blood oxygen parameter sequence The temperature sensor records the temperature change inside the mask cavity in real time, which is expressed as a temperature parameter sequence

[0038] Thus, a multimodal original physiological data set is constructed:

[0039]

[0040] Where t represents the sampling time and N represents the number of time steps in the current observation period.

[0041] S120 , performing timestamp alignment, missing value patching, and synchronous sampling processing on the multimodal original physiological data set to obtain a time-series aligned multimodal physiological data sequence.

[0042] Furthermore, for the multimodal original physiological data set D (1) , performing time series synchronization processing of multidimensional data. Due to the differences in the original sampling frequencies of various types of sensors, it is necessary to align the timestamps of each sequence, using linear interpolation or local regression interpolation (such as LOESS) to fill in the missing timestamp points to form a reconstructed sequence with equal time steps:

[0043]

[0044] where N ' To unify the number of time steps after resampling, it can usually be set to N ' =1000 steps / hour.

[0045] During the processing, for data segments with missing values ​​(such as If there is abnormal missing data in the interval, interpolation is performed based on the average value and change trend of adjacent intervals to ensure a stable equal-interval series.

[0046] S130 , performing channel normalization, feature extraction, and feature fusion on the multimodal original physiological data sequence obtained by processing in step S120 to generate a physiological feature sequence in a unified format.

[0047] After obtaining the multimodal raw physiological data sequence D after time alignment (2) After that, each sensor channel is further normalized to eliminate the unit dimension effect. The minimum and maximum normalization method is used to map each type of parameter to the interval [0,1]. The normalization formula is as follows:

[0048]

[0049] in: represents the value of the original j-th sensor at time t; is the standard value after normalization; min(X (2),j )、max(X (2),j ) represent the minimum and maximum values ​​of the parameter during the observation period.

[0050] After normalization, the key statistical features (mean, standard deviation, maximum rate of change) and frequency domain features (dominant frequency, Fourier transform amplitude, etc.) within each time window are extracted and recorded as a feature vector group:

[0051]

[0052] The function f(·) represents the feature extraction operation applied to a single sensor channel, and concat(·) represents the multi-dimensional feature concatenation operation to generate a physiological feature sequence in a unified format:

[0053]

[0054] Description of the connection and data transmission:

[0055] The physiological feature sequence F generated in step S130 (3) , is passed as input to the embedded neural network model in S210 for performing preliminary recognition of abnormal respiratory state. Specifically, the model will use F (3) As the input sample features, combined with the parameter weights obtained through training, time series classification and probability judgment are performed to generate respiratory behavior labels and abnormal probability parameters (see subsequent S210 description).

[0056] Step S200 at least includes steps S210-S230:

[0057] S210 , inputting the physiological feature sequence into a preset embedded neural network model, performing intelligent recognition on the respiratory rhythm and signal fluctuation pattern, and obtaining a first type of feature response.

[0058] Based on the physiological feature sequence generated in step S130 The physiological feature sequence is used as a time series input vector and input into a lightweight neural network model deployed on a local embedded chip.

[0059] Specifically, the neural network model is constructed using a gated recurrent unit (GRU) structure, and the input matrix is ​​defined as:

[0060]

[0061] in: represents a sliding window feature subsequence of length T starting from the i-th time step; T is the preset time window length, such as T = 30; is the feature vector normalized and fused in S130.

[0062] After the input sequence is input into the embedded GRU model, the first type of temporal response output is calculated The calculation formula is:

[0063]

[0064] in: is the hidden state output of the model in time window i; θ represents the set of weight parameters in the model, including reset gate weights, update gate weights and candidate state weights, which comes from the historical model parameter fitting during the training process.

[0065] described It represents the dynamic behavior response to the input data sequence in the time dimension and is subsequently used for respiratory event feature extraction.

[0066] S220. Based on the first type of characteristic response, extract target respiratory behavior related indicators, classify and calculate the respiratory cycle rate, abnormal peak position and signal stability, and obtain a respiratory behavior label.

[0067] Understandably, the first type of characteristic response sequence obtained in S210 Perform decoding and behavioral event recognition, and extract target respiratory indicators using multi-feature mapping, including:

[0068] Respiratory cycle rate index By calculating the period difference between the adjacent airflow rising boundaries, it can be expressed as:

[0069]

[0070] in: is the rising edge moment of the kth breath; δ is the minimum interval threshold for cycle recognition; Represents a Boolean conditional function (equals 1 if the condition is met); The unit is times / minute, which is subsequently used to classify and determine fast breathing, slow breathing or normal state.

[0071] Abnormal peak position identification It is obtained by detecting the position where the signal oscillation point exceeds the statistical fluctuation range.

[0072] Signal stability parameters By normalizing the standard deviation and the maximum variation within the sliding window, it can be expressed as:

[0073]

[0074] Finally, based on the above three characteristic parameters Call the predefined behavior classifier mapping function g(·) to generate the breathing behavior label:

[0075]

[0076] in

[0077] S230 , performing association modeling on the respiratory behavior labels, calculating respiratory abnormality probability value parameters, and forming a preliminary recognition output set.

[0078] Furthermore, a probability model of abnormal respiratory behavior is constructed to identify the behavior label in S220. Conduct time series correlation modeling. Introduce a weighted logic probability model, combine historical label change trends, and calculate anomaly probability values. It is defined as:

[0079]

[0080] in: Indicates whether the current label is abnormal; represents the frequency of label changes, that is, the number of behavioral state switches per unit time; α1, α2, and α3 are weighting coefficients, satisfying α1+α2+α3=1, and can be set to {0.5, 0.3, 0.2}, which are obtained from fitting historical patient data; Indicates the probability value of abnormal breathing state in the current period.

[0081] in, is the indicator function for abnormal respiratory status, which is defined as follows:

[0082]

[0083] in, is the respiratory behavior label generated in step S220, the classification result in the i-th time period, and the label set includes multiple behavioral states such as "normal breathing", "rapid breathing", "shallow breathing", and "intermittent breathing".

[0084] Finally, a preliminary recognition output set is generated:

[0085]

[0086] Wherein M is the total number of sliding windows, and the preliminary recognition output set will be used as the input of the S300 module for the subsequent data compression and reporting scheduling process.

[0087] Step S300 at least includes steps S310-S330:

[0088] S310: Based on the abnormal probability value parameter, the physiological feature sequence is divided into risk levels, and a corresponding compression encoding method is selected according to the risk level to generate compressed feature data.

[0089] Specifically, in combination with the abnormal probability value parameter sequence calculated in S230 above:

[0090]

[0091] Establishing risk level function The abnormal probability value of each period The risk levels are divided into the following three categories:

[0092]

[0093] Among them, τ1=0.3 and τ2=0.7 are risk thresholds set empirically, which are derived from the statistical results of multiple groups of clinical training samples.

[0094] According to the above risk level Select different compression strategies for physiological feature sequences Perform compression encoding. Compression methods include but are not limited to the following three categories:

[0095] Low risk level: A simplified downsampling method based on a mean window is used, with a compression rate of ρ L =80%;

[0096] Medium risk level: Principal component analysis (PCA) dimensionality reduction and compression method are used, with a compression rate of ρ M =60%;

[0097] High risk level: retain the original resolution and only perform lightweight lossless encoding with a compression ratio of ρ H =20%.

[0098] The compressed feature data is expressed as:

[0099]

[0100] Among them, the function Compress(·) is based on the risk level The compression algorithm selected is executed, and the output is the compressed feature data

[0101] S320: Setting the upload period and frequency of wireless communication according to the risk level and abnormal probability value parameters, and constructing an adaptive data scheduling strategy.

[0102] Furthermore, for each period of risk level and abnormal probability value parameters Building a dynamic dispatch function Indicates the upload cycle:

[0103]

[0104] in: Indicates the data upload period in the current cycle (unit: seconds); T max Indicates the maximum upload period limit, for example, it is set to 600 seconds; γ is the reduction factor, which is adjusted according to the system load, for example, γ = 500.

[0105] This formula is used to dynamically shorten the data upload cycle during high-risk periods, increase the communication frequency, and achieve data priority layered transmission.

[0106] Coordinate the upload cycle with compressed data Create a task schedule Used to perform specific communication tasks.

[0107] S330: Upload the compressed feature data and respiratory behavior tags to a remote cloud platform via a low-power wireless communication module.

[0108] Understandably, according to the constructed data scheduling table S (7) , calling a low-power wireless communication module (such as BLE or NB-IoT module) to complete the encapsulation and transmission of compressed feature data and respiratory behavior labels.

[0109] Specifically, the encapsulation format is:

[0110]

[0111] in: is the compressed feature data; is the breathing behavior label, which comes from S220 output; Data upload cycle; For risk level.

[0112] Under the premise of meeting the bandwidth and power consumption thresholds, each encapsulated data packet The data is sent to the remote cloud platform server via a low-power wireless communication module. The communication protocol can use MQTT or CoAP low-bandwidth protocols to ensure responsiveness and connection stability.

[0113] Description of the connection between the front and back:

[0114] The abnormal probability value parameter generated in step S230 It is the basic parameter for risk level classification in S310;

[0115] Compressed feature data in S310 Will be directly used for assembling and uploading S330 data packets;

[0116] The data upload cycle It is used to dynamically adjust the communication frequency and will interact with the server timing synchronization module in S400;

[0117] The minimum period for data upload depends on the system power budget and network load.

[0118] Step S400 at least includes steps S410-S430:

[0119] S410: Obtain the compressed feature data and respiratory behavior labels uploaded to the remote cloud platform, and construct a multi-dimensional input data set by combining historical health data and user portrait information.

[0120] In step S330, the portable smart breathing mask uploads a data packet to the cloud server via the wireless communication module. The data packet includes the compressed feature data. Breathing behavior tags Data upload cycle and risk level

[0121] At the same time, the cloud platform has stored the user's historical health data H (0) and user portrait data P (0) ,in:

[0122] Represents trend parameters such as the user's respiratory cycle rate and blood oxygen level over the past K days;

[0123] P (0)It is a static user portrait, including structured descriptive variables such as age, gender, underlying disease labels, allergy history, and previous hospitalization records.

[0124] Based on the above data sources, construct a multidimensional input dataset:

[0125]

[0126] Among them: the symbol ‖ represents the vector splicing operation; is the multidimensional input data set of the i-th observation period, that is, the joint data input vector, which serves as the input basis of the cloud-based analysis model.

[0127] S420: Input the multi-dimensional input data set into the cloud-based deep model, extract time series trend features and individual abnormal pattern features, and generate remote health assessment parameters.

[0128] Furthermore, the combined data is input into the vector Input is a deep time series analysis model built on the cloud. This model is based on a bidirectional gated recurrent network (BiGRU) and an attention mechanism. Its core calculation process is as follows:

[0129]

[0130] in: represents the extracted deep feature vector of the i-th time period; BiGRU(·) represents the bidirectional temporal modeling operation performed on the input data; Attention(·,ω) represents the weighted output after the introduction of the attention mechanism, where ω is the attention weight matrix, which is adaptively generated during training.

[0131] According to the depth feature vector Construct a remote health assessment parameter set:

[0132]

[0133] Where: μ r : expected value of the predicted respiratory cycle rate; σ r : standard deviation of predicted respiratory fluctuations; θ a : confidence level of the detected high-risk abnormal pattern; η c :With specific underlying disease type c∈P (0) The risk coupling degree represents the individualized correlation parameter.

[0134] The above parameters provide continuity input for subsequent trend tracking models.

[0135] S430: Based on the remote health assessment parameters, a trend tracking model is established to generate a user respiratory health risk trend sequence.

[0136] Furthermore, combining remote health assessment parameters within multiple time periods The data is input into the trend tracking model built in the cloud to calculate the evolution trend of the user's respiratory health risks.

[0137] Specifically, the exponentially weighted moving average (EWMA) method is used to establish the trend function:

[0138]

[0139] in: represents the risk trend value at the tth time step; θ a,t is the abnormal confidence of the current period; α r ∈(0,1) is the smoothing factor, for example, α r =0.6, based on fitting of past data; is the trend value of the previous period.

[0140] The final output is the user's respiratory health risk trend sequence in continuous time:

[0141]

[0142] Risk Trend Series It can be used for dynamic threshold setting and multi-level warning strategy generation in the subsequent module S500.

[0143] Instructions for connecting the previous and next steps:

[0144] described Both are derived from the previous steps S310 and S220;

[0145] The constructed multidimensional input vector It is the integration of front-end uploaded data and back-end historical data;

[0146] Obtained evaluation parameters is used as input to the trend function to form an individualized continuous risk evolution sequence And serves as the input basis for S500.

[0147] Step 500 at least includes steps S510-S530:

[0148] S510: Obtain the remote health assessment parameters and the user's respiratory health risk trend sequence, compare and analyze the current user status with the preset threshold of the historical model, and obtain a deviation parameter set.

[0149] The remote health assessment parameters generated in step S430 Trend sequence of respiratory health risks of users Serves as the core evaluation basis for the current user status.

[0150] At the same time, the system has a static multi-level security threshold model preset:

[0151]

[0152] Perform difference calculation on each indicator and construct the deviation parameter set Δ (10) :

[0153]

[0154] The set of bias parameters is defined as:

[0155]

[0156] This set is used to quantify the degree of deviation between the current analysis parameters and the historical safety range, and is the basis for subsequent dynamic adjustments.

[0157] S520: Based on the deviation parameter set, adjust the trigger conditions in the multi-level threshold judgment model to construct a dynamically updated dynamic threshold policy set.

[0158] Furthermore, according to the deviation parameter set Δ calculated in step S510 (10) , a weighted strategy is used to generate a dynamic threshold function:

[0159]

[0160] in: Represents the dynamic threshold parameter set in the current cycle i; α1, α2, α3, α4 are model weights, satisfying Its value is derived from the historical sample fitting results, for example: α1 = 0.3, α2 = 0.3, α3 = 0.2, α4 = 0.2;

[0161] The weighting strategy is used to emphasize the impact of the current health parameter deviation on the dynamic threshold, solving the problem of poor adaptability of static threshold to individual users.

[0162] Finally, the dynamically updated dynamic threshold policy set is formed:

[0163]

[0164] Used for subsequent trigger condition determination process.

[0165] S530: Input the latest remote health assessment parameters into the dynamic threshold policy set for judgment, and generate corresponding alarm data and response level labels.

[0166] Based on the remote health analysis parameters output in step S420 Combine it with the dynamic threshold parameter Compare item by item and generate multi-level alerts according to preset strategies.

[0167] Set the risk level determination function:

[0168]

[0169] in: is the response level label of the i-th cycle;

[0170] The response level is driven by the dual indicators of the abnormality degree and trend value of the evaluation parameters;

[0171] This strategy integrates sudden anomalies (such as θ a surge) and continuous risk accumulation (such as The rising) factor reflects the intelligent alarm mechanism with high time sensitivity.

[0172] Finally, structured alert data is generated:

[0173]

[0174] described It will serve as the input for the visual display and feedback record in the subsequent module S600.

[0175] Instructions for connecting the previous and next steps:

[0176] Remote evaluation of input parameters and risk trends Derived from S420 and S430;

[0177] Deviation parameter set Δ (10) Used to dynamically adjust the threshold parameter set Establish an individualized judgment mechanism;

[0178] Output alarm data With responsive grade labels Provide basic information for subsequent remote feedback interaction.

[0179] Step S600 at least includes steps S610-S630:

[0180] S610 obtains the alarm data and the user's respiratory health risk trend sequence, structures and organizes them, and processes them graphically to generate remote visualization display content.

[0181] Specifically, the alarm data set is obtained from step S530:

[0182]

[0183] in: assessing parameters for telehealth; Breathe health risk trend sequence for users; is a set of dynamic threshold parameters; is the response level label.

[0184] Based on this structured data, we organize the visualization parameters and construct the following graphical display indicator set:

[0185]

[0186] in: Indicates the ratio of respiratory cycle rate to safety benchmark; Represents the normalized index of respiratory fluctuation level; represents the abnormal confidence ratio; It is the risk trend value, which is used directly in the drawing.

[0187] Map ratio indicators to various visual graphic components such as radar charts, line charts, color risk charts, etc., and output them in a structured manner to the user-side application interface.

[0188] S620: Display the visual display content on the user terminal, provide a multi-dimensional interactive input interface, and receive the user's subjective state feedback data and auxiliary health record data uploaded by the user terminal.

[0189] Furthermore, the remote visual display content generated in step S610 is displayed on the user terminal device (such as mobile phone APP, smart watch interface) The user interaction input interface is loaded synchronously to collect the following two types of feedback data:

[0190] Subjective state feedback data The user actively selects or fills in the current physiological status, including but not limited to:

[0191] Feeling of shortness of breath (yes / no);

[0192] Chest tightness / cough / dizziness (multiple choices);

[0193] Sleep status score (0-10);

[0194] Pain score (0-10), etc.

[0195] Auxiliary health record data Auxiliary records uploaded synchronously by users or the system, including:

[0196] Medication records;

[0197] amount of exercise;

[0198] Dietary status;

[0199] Sleep time, etc.

[0200] The feedback data is encapsulated into a joint feedback data structure through the interactive panel:

[0201]

[0202] Where T i Record timestamps for feedback to facilitate matching with remote server-side data.

[0203] S630: Standardize the user's subjective state feedback data and the auxiliary health record data to form a structured interactive health feedback data set.

[0204] Specifically, the above joint feedback data structure The input standardization processing module normalizes the format of the feedback content, maps variables, and fills in missing information.

[0205] Text / selection format mapping: For example, "chest tightness" is mapped to an indicator dimension variable Otherwise it is 0;

[0206] Normalization of score items: Sleep scores are normalized to the interval [0,1], expressed as

[0207] Behavioral data classification: auxiliary records are organized into multi-dimensional time series matrices by category, such as Indicates the value of the j-th record.

[0208] The structured interactive health feedback dataset is constructed as follows:

[0209]

[0210] Finally, the structured feedback dataset Compared with the aforementioned server-side data Realize joint modeling input and provide basic support for the next step of the S700 module to implement policy feedback-driven parameter adaptation.

[0211] Description of the connection between the front and back:

[0212] Derived from S530, it is the input for visual display;

[0213] Risk Trend Series Provided by S430, it is the core data for chart drawing;

[0214] It is the output of S600 and will be passed to the S700 module as user feedback data.

[0215] Step S700 at least includes steps S710-S730:

[0216] S710 : Based on the remote health assessment parameters and the interactive health feedback data set, update the behavior recognition parameters and the abnormal classification threshold parameters.

[0217] Specifically, obtain the remote health assessment parameters output in step S420 And the structured interactive health feedback dataset generated in step S630:

[0218]

[0219] According to the system's built-in update function, the parameters of the respiratory behavior recognition model deployed on the edge are fine-tuned, and the recognition weight vector in the model is updated using a weighted fusion strategy. Its expression is:

[0220]

[0221] Where: β1 and β2 are fusion weight coefficients, satisfying β1+β2=1. For example, β1=0.6 and β2=0.4 can be taken, which are derived from offline optimization experiments.

[0222] Represents the set of updated weights for multi-channel feature maps within a time window in the edge model.

[0223] Furthermore, based on the new parameters Reconstructing anomaly classification threshold set The specific definitions are as follows:

[0224]

[0225] in: Critical threshold of respiratory cycle changes; Blood oxygen drop fluctuation threshold; Abnormal pattern detection probability threshold.

[0226] This threshold set will be used to update the edge model's judgment basis for newly collected data.

[0227] S720: Input the updated model parameters into the front-end control system, dynamically configure the sampling frequency, power consumption level, and inference cycle of the sensor, and generate a local configuration instruction set.

[0228] Understandably, the update weight set generated in step S710 With threshold set Input into the local embedded control system to build the sensor operation parameter mapping strategy.

[0229] First, the sensor sampling frequency is dynamically calculated based on the risk assessment level and the confidence level of the behavior model.

[0230]

[0231] Where: f base is the default sampling frequency (such as 1Hz); δ f is the frequency increase step size, for example, 0.5Hz; θ a Originated from is the current abnormal probability judgment threshold.

[0232] Then, combine the device power budget with the response level Setting energy consumption levels and the inference cycle And encapsulate the three into a set of local configuration instructions:

[0233]

[0234] S730: Call the local configuration instruction set to complete the dynamic regulation of the front-end data acquisition process and the embedded model reasoning cycle.

[0235] Furthermore, the local control unit automatically calls the configuration instruction set Perform dynamic control of the following modules:

[0236] Sensor module control: according to Set the sampling rate of airflow, humidity, blood oxygen and temperature channels;

[0237] Power consumption module control: according to Control chip operating frequency and sleep state switching strategy;

[0238] Reasoning cycle control: According to Control the edge model running frequency (i.e., perform model inference every few seconds).

[0239] Ultimately, the edge system updates parameters based on remote health assessment parameters and interactive health feedback data sets, and adapts to current usage scenarios and risk levels in real time.

[0240] Description of the connection between the front and back:

[0241] The input of this module comes from S420 S630

[0242] Update results Driver model and threshold updates;

[0243] Output Configuration Collection It is a parameter entity that directly acts on the local hardware system control layer;

[0244] The system feedback loop is closed here.

[0245] The present invention effectively enhances the synergy and temporal stability between data features and improves the accuracy of model recognition by uniformly aligning and weighting the mutual information of multimodal raw physiological data such as airflow, humidity, blood oxygen saturation and temperature; it uses an embedded neural network model to realize the preliminary recognition of respiratory behavior at the edge, with low latency and local reasoning capabilities, reducing dependence on network continuity; the present invention introduces an adaptive upload strategy based on abnormal probability value parameters for the first time. Through the above strategy, high-frequency upload is achieved under high-risk conditions, and the upload cycle is extended under low-risk conditions, taking into account both communication energy consumption and real-time performance; the cloud model constructed by combining BiGRU and attention mechanism performs personalized modeling analysis on uploaded data and user portraits to generate risk trend sequences; and through a dynamic multi-level early warning strategy, the local model parameters and sampling strategy are automatically updated in combination with user feedback data, realizing adaptive control of data acquisition frequency and model reasoning cycle. The above method effectively improves the data processing efficiency, risk identification accuracy and remote health management capabilities of the equipment, and is suitable for respiratory health monitoring in chronic disease populations, the elderly and special scenarios.

[0246] Example 2: Figure 2 FIG. 1 shows a structural block diagram of a portable intelligent breathing mask and a remote monitoring system thereof according to an embodiment of the present invention. Figure 2 As shown, the structure may include:

[0247] The sensor acquisition module 10 is used to acquire multimodal physiological signal data generated by the user while wearing a respiratory mask. Specifically, this includes, but is not limited to, airflow parameters, humidity parameters, blood oxygen saturation parameters, and cavity temperature parameters. By integrating a miniature airflow sensor, humidity sensor, blood oxygen saturation sensor, and temperature sensor within the respiratory mask, this data is continuously acquired, forming an initial time-series raw data set.

[0248] The edge processing and preliminary judgment module 20 is used to synchronize timestamps, repair missing data, normalize channels, extract features, and perform fusion processing on the raw time series data to construct a physiological feature sequence in a unified format. Furthermore, the module is embedded with a lightweight time series neural network model (such as a GRU or LSTM structure), which receives the feature sequence input, performs local preliminary recognition of patterns such as respiratory rhythm and fluctuation anomalies, and outputs a respiratory behavior label and the corresponding abnormality probability value.

[0249] The adaptive data compression and communication scheduling module 30 is used to classify data risk levels based on the aforementioned abnormal probability value parameters and select compression coding strategies (such as PCA compression and downsampling coding) for different risk levels to encode and compress the physiological feature sequences. Furthermore, the module sets different data upload cycles and communication frequencies based on the current risk level, and uploads the compressed feature data and tag information to a remote cloud platform via low-power wireless communication protocols such as BLE or NB-IoT.

[0250] The cloud-based modeling and risk analysis module 40 receives compressed feature data and respiratory behavior labels from the front-end device, combines historical health data with user profile data, and constructs a multi-dimensional analysis input vector. Based on a deep time series analysis model (such as BiGRU combined with an attention mechanism), this module extracts temporal trend characteristics of individual respiratory health and individualized abnormal pattern characteristics, and generates remote health assessment parameters and respiratory risk trend sequences.

[0251] The multi-level warning determination module 50 is used to construct a dynamic threshold determination model based on the remote health assessment parameters and risk trend sequence. This module adjusts the deviation of historical static thresholds to generate a multi-level dynamic warning strategy set for the current cycle. It then determines whether to trigger an alarm based on the user's anomaly recognition confidence and risk trend value, and outputs the corresponding response level label (e.g., Level 1 Severe, Level 2 Moderate, Level 3 Observation, etc.).

[0252] The remote visualization interaction module 60 is used to graphically display the health assessment results, risk trend charts, warning levels, and other content output by the system to the user terminal (such as a mobile app, wearable device, etc.). This module allows users to input subjective status feedback (such as dyspnea, perceived discomfort, medication use, etc.) and auxiliary health record data based on the displayed results, and upload them to the cloud for subsequent model correction.

[0253] The local policy feedback control module 70 is used to receive remote health assessment parameters and interactive health feedback datasets generated by the cloud, update the behavior recognition parameters and anomaly classification threshold parameters in the local edge recognition model, and generate a set of local control instructions, including sensor sampling frequency, power consumption level, and inference cycle. Furthermore, this module controls the front-end embedded system to adjust the data collection strategy and model execution frequency in real time, achieving intelligent and adaptive operation of the mask device.

[0254] The portable intelligent breathing mask and its remote monitoring system provided by the embodiments of the present invention have the following beneficial effects:

[0255] Integrate multimodal high-frequency physiological signal acquisition to achieve accurate dynamic perception of respiratory status;

[0256] Build an edge-cloud collaborative intelligent recognition mechanism to achieve graded prediction and early warning response of abnormal respiratory behavior;

[0257] Introducing a risk assessment model based on individual health profiles to improve the personalization and accuracy of respiratory health management;

[0258] Establish a closed-loop parameter adaptation mechanism driven by user feedback to significantly enhance the device's adaptability to different users and scenarios;

[0259] It achieves low-power communication and operation strategy optimization for devices, suitable for mobile environments and long-term usage scenarios.

[0260] Obviously, the embodiments described above are only some of the embodiments of the present application, rather than all of the embodiments. The preferred embodiments of the present application are given in the accompanying drawings, but they do not limit the patent scope of the present application. The present application can be implemented in many different forms. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present application more thorough and comprehensive. Although the present application has been described in detail with reference to the aforementioned embodiments, for those skilled in the art, it is still possible to modify the technical solutions described in the aforementioned specific embodiments, or to make equivalent replacements for some of the technical features therein. Any equivalent structure made using the contents of the present application specification and the accompanying drawings, directly or indirectly used in other related technical fields, is also within the scope of patent protection of the present application.

Claims

1. A remote monitoring method for a portable intelligent breathing mask, characterized in that: The method comprises: Acquiring multimodal raw physiological data generated by the user while wearing the mask, the raw physiological data including airflow parameters, humidity parameters, blood oxygen parameters, and temperature parameters, and performing timestamp alignment and feature fusion on the multimodal raw physiological data to obtain a physiological feature sequence in a unified format; Inputting the physiological feature sequence into an embedded neural network model to perform respiratory behavior recognition and obtain a respiratory behavior label and an abnormal probability value parameter; Based on the abnormal probability value parameter, the physiological feature sequence is compressed and encoded to obtain compressed feature data, and a data upload cycle is set: The compressed feature data and the respiratory behavior label are uploaded to the cloud platform via a low-power wireless communication module, and then input into a deep model built based on a bidirectional gated recurrent unit and an attention mechanism to generate a remote health assessment parameter and a sequence of the user's respiratory health risk trend; generating alarm data and response level labels based on the remote health assessment parameters and the user's respiratory health risk trend sequence; Based on the remote health assessment parameters and the interactive health feedback data set, the edge model and sensor sampling strategy are updated to achieve dynamic regulation of the front-end acquisition frequency and inference cycle.

2. The remote monitoring method according to claim 1, wherein: The feature fusion includes channel normalization, feature extraction and feature fusion of the airflow parameters, humidity parameters, blood oxygen parameters and temperature parameters.

3. The remote monitoring method according to claim 1, wherein: The embedded neural network model is a gated recurrent unit (GRU) and / or a long short-term memory (LSTM) network structure.

4. The remote monitoring method according to claim 1, wherein: The compression coding process includes selecting at least one compression method of mean downsampling, principal component analysis (PCA) and / or lightweight lossless compression based on the abnormal probability value parameter.

5. The remote monitoring method according to claim 1, wherein: The low-power communication module includes any one of a Bluetooth BLE module, an NB-IoT module and / or a LoRa module.

6. The remote monitoring method according to claim 1, wherein: The deep model is a temporal modeling network constructed based on a bidirectional gated recurrent unit BiGRU and an attention mechanism.

7. The remote monitoring method according to claim 1, wherein: The remote health assessment parameters include respiratory cycle expected value, fluctuation standard deviation, abnormality confidence and individual portrait correlation weight.

8. The remote monitoring method according to claim 1, wherein: The alarm data is generated by weighted adjustment of static thresholds based on remote health assessment parameters and user respiratory health risk trend values, supporting three-level response level judgment.

9. The remote monitoring method according to claim 1, wherein: The interactive health feedback dataset is obtained based on user subjective state feedback data and auxiliary health record data.

10. The remote monitoring method according to claim 1, wherein: The sensor sampling strategy includes sampling frequency, power consumption level and inference cycle configuration, and the dynamic regulation is driven jointly based on the remote health assessment parameters and the interactive health feedback dataset.

Images