Intelligent control method and system for intermittent jet ventilation nasal oxygen inhalation catheter

Abstract

The invention provides an intelligent control method and system for an intermittent jet ventilation nasal oxygen inhalation catheter, and relates to the technical field of intelligent control, and the method comprises the steps: 1, inputting the data of a multi-modal sensor into a self-adaptive neural network of an embedded gating circulation unit, and carrying out the feature extraction and state memory, so as to obtain a feature vector; 2, mapping the feature vectors to a high-dimensional physiological state space, and determining state reference coordinates based on the feature vectors of the blood oxygen saturation and the respiratory frequency; and calculating an evolution direction difference between the blood oxygen saturation feature vector and the respiratory frequency feature vector, and dividing a state boundary. Through multi-mode sensing, state mapping, dynamic correction, error feedback and motion sudden change pre-judgment, real-time adaptation of intermittent jet ventilation parameters of the nasal oxygen inhalation catheter in the plateau mountaineering process is achieved, it is guaranteed that oxygen output is dynamically matched with physiological needs of a user, and the oxygen supply efficiency is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent control, in particular to an intermittent jet ventilation nasal oxygen catheter intelligent control method and system. BACKGROUND

[0002] In the process of high-altitude mountaineering activities, the user often needs to assist oxygen supply through the nasal oxygen catheter. The intelligent control method in the prior art mostly relies on a single signal such as blood oxygen saturation or respiratory rate for adjustment. Such a scheme can generally meet the basic oxygen supply demand to a certain extent when the user is in a steady-state activity (such as gentle walking or short rest). However, when the user suddenly changes from a resting state to a dynamic scene such as climbing a steep slope, the physiological state will change rapidly and continuously. At this time, the traditional control method is often difficult to fully integrate multi-modal physiological parameters and environmental information, and there may be certain room for improvement in the identification accuracy and timeliness of dynamic demand.

[0003] Some methods using traditional neural network structures have certain signal processing capabilities, but when faced with long-term and strong dynamic changes in the high-altitude mountaineering scene, their memory and association capabilities for historical states may be limited, making it difficult to accurately capture the coupling rules and evolution trends between respiratory and blood oxygen signals. For example, when the user's exercise intensity suddenly increases, due to insufficient learning and use of physiological sequence characteristics before the state changes, the system may have a prediction delay for the possible downward trend of blood oxygen saturation, resulting in a lag in the generation of adjustment instructions, and in some cases, the oxygen supply does not fully match the actual demand. SUMMARY

[0004] The technical problem to be solved by the present application is to provide an intermittent jet ventilation nasal oxygen catheter intelligent control method and system to realize accurate matching of oxygen output and real-time physiological demand.

[0005] To solve the above technical problems, the technical solution of the present application is as follows: In a first aspect, an intermittent jet ventilation nasal oxygen catheter intelligent control method is provided, the method comprising: Step 1: inputting multi-modal sensor data into an adaptive neural network embedded in a gated recurrent unit for feature extraction and state memory to obtain a feature vector; Step 2: mapping the feature vector to a high-dimensional physiological state space, determining a state reference coordinate based on the blood oxygen saturation and respiratory rate feature vectors, calculating the evolution direction difference between the blood oxygen saturation feature vector and the respiratory rate feature vector, and dividing the state boundary; Step 3: setting four state feature points within the state boundary, generating a correction factor based on the spatial distribution characteristics of the state feature points, correcting the feature vector using the correction factor, and obtaining a corrected feature vector; Step 4, the corrected feature vector is calculated to generate a control instruction set containing target injection frequency, driving pressure and inspiration phase time ratio; the piezoelectric ceramic high-frequency valve is driven to perform intermittent oxygen injection according to the control instruction set, and actual gas flow and pressure data are obtained through the sensor, and the actual gas flow and pressure data are compared with the target control instruction set to generate an error signal feedback to the neural network to adjust the network parameters; Step 5, when a sudden change in user motion intensity is detected, the blood oxygen saturation change trend in the future preset time is predicted according to the physiological state historical data memorized by the gated recurrent unit, and the generation strategy of the control instruction set is adjusted to realize the matching of oxygen output and real-time physiological demand.

[0006] The second aspect is an intelligent control system for a nasal oxygen catheter capable of intermittent injection of ventilation, comprising: The extraction module is used for embedding multi-modal sensor data into an adaptive neural network of a gated recurrent unit for feature extraction and state memory, so as to obtain a feature vector; The projection module is used for mapping the feature vector to a high-dimensional physiological state space, determining a state reference coordinate based on the blood oxygen saturation and the respiratory rate feature vector, calculating the evolution direction difference of the blood oxygen saturation feature vector and the respiratory rate feature vector, and dividing a state boundary; The correction module is used for setting four state feature points within the state boundary, generating a correction factor based on the spatial distribution characteristics of the state feature points, and correcting the feature vector by using the correction factor to obtain a corrected feature vector; The feedback module is used for calculating the corrected feature vector to generate a control instruction set containing target injection frequency, driving pressure and inspiration phase time ratio; the piezoelectric ceramic high-frequency valve is driven to perform intermittent oxygen injection according to the control instruction set, and actual gas flow and pressure data are obtained through the sensor, and the actual gas flow and pressure data are compared with the target control instruction set to generate an error signal feedback to the neural network to adjust the network parameters; The mutation module is used for predicting the blood oxygen saturation change trend in the future preset time according to the physiological state historical data memorized by the gated recurrent unit when a sudden change in user motion intensity is detected, and adjusting the generation strategy of the control instruction set to realize the matching of oxygen output and real-time physiological demand.

[0007] The third aspect is a computing device, comprising: one or more processors; a storage device for storing one or more programs, when the one or more programs are executed by the one or more processors, so that the one or more processors implement the method.

[0008] In a fourth aspect, a computer-readable storage medium stores a program, which implements the method when executed by a processor.

[0009] The above solution of the present invention includes at least the following beneficial effects: By processing multimodal sensor data through an adaptive neural network embedded in a gated recurrent unit, the system can not only integrate physiological signals such as blood oxygen, respiration, and electrocardiogram (ECG) with environmental parameters such as air pressure and temperature, but also effectively memorize the temporal variations of physiological states. This solves the problem of relying on a single signal and difficulty capturing the coupling patterns between respiration and blood oxygen, thereby improving the comprehensiveness of physiological state identification. A high-dimensional physiological state space mapping and state boundary demarcation mechanism, combined with correction factors generated from state feature points, can accurately quantify the evolution differences between blood oxygen and respiratory rate, reducing the bias of single parameter judgment and making the feature vector more consistent with the actual physiological state. A neural network parameter adjustment mechanism based on error feedback dynamically optimizes model output by comparing actual gas parameters with target instructions in real time, helping to narrow the deviation between oxygen output and theoretical demand. A predictive adjustment strategy for sudden changes in exercise intensity leverages the gated recurrent unit's ability to memorize historical data to predict blood oxygen trend changes and adjust the control strategy in advance. This reduces response lag during rapid physiological changes and ensures that oxygen output more promptly matches the user's real-time needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 The present invention provides a flow chart of an intelligent control method for a nasal oxygen inhalation catheter capable of intermittent jet ventilation.

[0011] Figure 2 This is a schematic diagram of an intelligent control system for a nasal oxygen inhalation catheter capable of intermittent jet ventilation provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0012] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0013] like Figure 1 As shown, an embodiment of the present invention provides an intelligent control method for a nasal oxygen inhalation catheter capable of intermittent jet ventilation, the method comprising the following steps: Step 1: Input the multimodal sensor data into the adaptive neural network embedded in the gated recurrent unit for feature extraction and state memory to obtain the feature vector; Step 2, mapping the feature vector to a high-dimensional physiological state space, determining a state reference coordinate based on the blood oxygen saturation and respiratory rate feature vectors; calculating the evolution difference between the blood oxygen saturation feature vector and the respiratory rate feature vector differences and divide state boundaries; Step 3, setting four state feature points within the state boundary, generating a correction factor based on the spatial distribution characteristics of the state feature points, and correcting the feature vector using the correction factor to obtain a corrected feature vector; Step 4, calculating the corrected feature vector to generate a control instruction set containing target injection frequency, driving pressure and inspiration-expiration phase time ratio; driving the piezoelectric ceramic high-frequency valve to perform intermittent oxygen injection according to the control instruction set, and obtaining actual gas flow and pressure data through the sensor, comparing the actual gas flow and pressure data with the target control instruction set to generate an error signal feedback to the neural network to adjust the network parameters; Step 5, when a sudden change in user motion intensity is detected, predicting the blood oxygen saturation change trend in the future preset time according to the physiological state historical data remembered by the gated recurrent unit, and adjusting the generation strategy of the control instruction set to realize the matching of oxygen output and real-time physiological demand.

[0014] In the embodiment of the present application, the adaptive neural network embedded with the gated recurrent unit processes multi-modal sensor data, which can not only fuse physiological signals such as blood oxygen, respiration, electrocardiogram and environmental parameters such as air pressure, temperature, but also effectively remember the time sequence change characteristics of physiological state, solve the problem of relying on a single signal and being difficult to capture the coupling rule of respiration and blood oxygen, and improve the comprehensiveness of physiological state recognition; the high-dimensional physiological state space mapping and state boundary division mechanism, combined with the correction factor generated by the state feature points, can accurately quantify the evolution difference between blood oxygen and respiratory rate, reduce the deviation of single parameter judgment, and make the feature vector more consistent with the actual physiological state; the neural network parameter adjustment mechanism based on error feedback compares the actual gas parameters with the target instructions in real time, dynamically optimizes the model output, which helps to reduce the deviation between oxygen supply output and theoretical demand; the prediction adjustment strategy for sudden change in motion intensity uses the memory capability of the gated recurrent unit to predict the blood oxygen change trend in advance and adjust the control strategy, which can reduce the response lag when the physiological state changes rapidly, and make the oxygen output more timely match the real-time demand of the user.

[0015] In another preferred embodiment of the present application, the process of acquiring multi-modal sensor data is as follows: In the embodiment of the present application, the physiological signals of the user and the highland environment parameters are collected in real time by the multi-modal sensor, the physiological signals include blood oxygen saturation, respiratory waveform, electrocardiogram signal and three-axis motion acceleration, and the environment parameters include atmospheric pressure, environmental temperature and airflow speed, the physiological signals of the user and the highland environment parameters are synchronized and time-aligned to obtain multi-modal sensor data, specifically including: integrating sensor modules on the nasal oxygen catheter body and the matching wearable device, wherein the blood oxygen sensor is attached to the finger or the ear lobe mucosa to monitor the blood oxygen saturation, the respiratory flow sensor is embedded in the catheter passage to collect the respiratory waveform, the electrocardiogram electrode patch is attached to the chest to obtain the electrocardiogram signal, and the three-axis acceleration sensor is integrated in the wearable device to detect the limb motion state; at the same time, the atmospheric pressure sensor, the temperature and humidity integrated sensor and the miniature airflow meter are installed at the air inlet of the catheter to collect the atmospheric pressure, the environmental temperature and the airflow speed respectively; the sensors are synchronously triggered by a unified clock source to collect data in real time at a sampling frequency of 50Hz, and each frame of data is attached with an accurate time stamp; the collected original signals are preprocessed, wherein the physiological signals are filtered by a band-pass filter to remove electromyographic interference and baseline drift, and the environment parameters are subjected to outlier rejection; the physiological data such as blood oxygen saturation, respiratory waveform, electrocardiogram signal and three-axis motion acceleration within the same time period are matched with the environment parameters such as atmospheric pressure, environmental temperature and airflow speed at the corresponding time point point by point according to the time sequence, and integrated into a multi-modal sensor data set containing 7 types of parameters.

[0016] In a preferred embodiment of the present application, step 1 comprises: In the embodiment of the present application, step 100, the multi-modal sensor data is inputted into the adaptive neural network embedded with the gated recurrent unit according to the time sequence, the time correlation between the blood oxygen saturation data, the respiratory waveform data, the electrocardiogram signal data and the three-axis motion acceleration data is established by the gated recurrent unit inside the adaptive neural network, and the time sequence change characteristics of the physiological signals are stored, specifically including: taking 30 seconds of continuous data collection in highland mountaineering as a time window, sorting the blood oxygen saturation data, the respiratory waveform data, the electrocardiogram signal data and the three-axis motion acceleration data in the window according to the collection time to form four-dimensional time sequence data, each dimension corresponding to one physiological signal; each data point is attached with a time stamp, the sampling frequency is 50Hz, there are 1500 data points in 30 seconds, and the time stamps 1-1500 correspond in turn; the four-dimensional time sequence data is inputted into the network, the network is first standardized to unify the numerical range of single physiological signal to 0-1, wherein the numerical value of each data point of single physiological signal is the minimum value in the 30-second window of the signal, and then ÷ (the maximum value minus the minimum value in the 30-second window of the signal); the blood oxygen saturation data, the respiratory waveform, the electrocardiogram signal and the three-axis motion acceleration data are calculated in the same way, and finally the standardized four-dimensional time sequence data is obtained.

[0017] The gating cycle unit is started, and the importance of the four-dimensional physiological signal at the current time point to the current state is judged by updating the gate, wherein the current time stamp standardized four-dimensional physiological signal data and the previous time stamp hidden state data (the hidden state is a 32-dimensional zero vector at the initial time stamp 0, and the dimension is set according to the model training requirement) are spliced to form the update gate splicing data; the update gate splicing data is input into a sigmoid activation function, and the formula is wherein is the update gate linear transformation result, and the calculation method is update gate splicing data x update gate weight matrix + update gate bias vector; e is a natural constant (about equal to 2.718), and the update gate weight (range 0-1) is obtained; the update gate weight matrix has a dimension of 4x32 (4 corresponds to the four-dimensional physiological signal, and 32 corresponds to the hidden state dimension), such as 0.12 in the first row and the first column, -0.08 in the second row and the third column, 0.09 in the third row and the fifth column, and -0.11 in the fourth row and the tenth column; the update gate bias vector has a dimension of 32, and all elements are 0.05 (adjusting the offset of the linear transformation result); the current time stamp standardized four-dimensional physiological signal data and the previous time stamp hidden state data are multiplied by the update gate weight respectively to obtain two intermediate results, and the two results are added to obtain the current time stamp candidate hidden state (temporarily stored current state feature); the current time stamp standardized four-dimensional physiological signal data and the previous time stamp hidden state data are spliced by the reset gate to form the reset gate splicing data (combining the two types of data into a continuous vector in order); the reset gate splicing data is input into a sigmoid activation function, and the formula is wherein is the reset gate linear transformation result, and the calculation method is reset gate splicing data x reset gate weight matrix + reset gate bias vector, to obtain the reset gate weight (range 0-1); the reset gate weight matrix has a dimension of 4x32, such as 0.10 in the first row and the second column, -0.07 in the second row and the fourth column, 0.08 in the third row and the sixth column, and -0.10 in the fourth row and the twelfth column; the reset gate bias vector has a dimension of 32, and all elements are 0.03 (adjusting the offset of the linear transformation result); the previous time stamp hidden state data is multiplied by the reset gate weight to obtain the adjusted previous time stamp hidden state data (filtered historical state feature); the adjusted previous time stamp hidden state data and the current time stamp standardized four-dimensional physiological signal data are spliced to form the hidden state splicing data (combining the two types of data into a continuous vector in order); the hidden state splicing data is input into a tanh activation function, and the formula is wherein is the candidate hidden state linear transformation result, and the calculation method is The hidden state splicing data candidate hidden state weight matrix + candidate hidden state bias vector, to obtain the current timestamp hidden state data (a feature vector fused with historical and current information, ranging from -1 to 1); the candidate hidden state weight matrix has a dimension of 4x32, such as 0.11 in the first row and the third column, -0.09 in the second row and the fifth column, 0.07 in the third row and the seventh column, and -0.08 in the fourth row and the eleventh column; the candidate hidden state bias vector has a dimension of 32, and all elements are 0.01 (adjusting the offset of the linear transformation result); the hidden state data contains time correlation of four types of data within a 30-second window, and the gated recurrent unit stores the data as a physiological signal time sequence change feature.

[0018] In step 101, the time sequence change feature is input into a fully connected neural network with multiple hidden layers in an adaptive neural network, and is transformed by a nonlinear activation function of multiple hidden layers to extract deep features reflecting dynamic changes of physiological load and environmental interference features, specifically including: a 32-dimensional time sequence change feature vector output by the gated recurrent unit in step 100 is called as input data of the first hidden layer of the fully connected neural network; the layer contains 64 neurons; for a single neuron, multiply each element of the 32-dimensional input vector with the corresponding 32-dimensional connection weight of the neuron to obtain 32 product results; after summing the 32 product results, add the bias value of the neuron to output the initial output value of the neuron, i.e., the first layer linear transformation result z4; all neuron connection weights are obtained by training 500 groups of highland physiological data sets, for example, the 5th column element of the 32-dimensional connection weight of the 1st neuron is 0.15, and the 10th column element is -0.12; the 12th column element of the 32-dimensional connection weight of the 20th neuron is 0.08, and the 25th column element is -0.09; the neuron bias value ranges from -0.1 to 0.1, for example, the 1st is 0.04, the 2nd is -0.02, and the 3rd is 0.05; the initial output value z4 of each neuron is input into a ReLU activation function, , the part greater than 0 is retained, and the part less than or equal to 0 is discarded; the results of 64 neurons after activation processing are combined into a 64-dimensional first layer output feature vector.

[0019] The 64-dimensional feature vector output by the first layer of hidden layers is input to the second layer of hidden layers containing 32 neurons. The single neuron operation logic is consistent with the first layer. Each element of the 64-dimensional input vector is multiplied by the corresponding 64-dimensional connection weight of the neuron to obtain 64 product results. After performing summation operation on the 64 product results, the result is added to the bias value of the neuron to output the initial output value of the neuron (i.e. the second layer linear transformation result z5). For example, in the 64-dimensional connection weight of the 5th neuron, the 10th column element is 0.11 and the 40th column element is -0.14. In the 64-dimensional connection weight of the 15th neuron, the 20th column element is 0.07 and the 50th column element is -0.08. The neuron bias value ranges from -0.1 to 0.1, for example, the 1st is 0.02, the 2nd is -0.03, and the 3rd is 0.04. The initial output value z5 of each neuron is input to the ReLU activation function, which retains the part greater than 0 and discards the part less than or equal to 0. The results of the 32 neurons after activation processing are combined into a 32-dimensional second layer intermediate output feature vector. The high-altitude environment parameters (atmospheric pressure, environmental temperature, air flow speed) are processed according to the physiological signal standardization logic of step 100. For a single environment parameter, the value of each data point in a 30-second time window is subtracted from the minimum value of the parameter in the window, and then divided by (the maximum value minus the minimum value of the parameter in the window) to obtain the standardized single environment parameter data. After standardizing the three types of parameters, a 3-dimensional standardized environment parameter vector is obtained. The correlation operation is performed on the standardized environment parameter vector and the second layer intermediate output feature vector, i.e. the first three elements of the standardized vector are multiplied by the first three elements of the second layer intermediate vector, and the fourth to thirty-second elements of the second layer intermediate vector are multiplied by 1 (the value remains unchanged). The above operation results together form a 32-dimensional environment interference correlation data. The 32-dimensional environment interference correlation data and the corresponding elements of the 32-dimensional second layer intermediate output feature vector are added to obtain a 32-dimensional second layer final output feature vector. Two types of features are separated from the 32-dimensional second layer final output feature vector, i.e. the first 24 elements are depth features (reflecting the dynamic changes of physiological load caused by changes in exercise intensity and hypoxia state during high-altitude mountaineering); the last 8 elements are environment interference features (reflecting the influence of atmospheric pressure fluctuations, temperature changes, and air flow interference on physiological signals in the high-altitude environment).

[0020] Step 102, the depth feature and the environmental interference feature are fused through a feature splicing operation to obtain a feature vector representing the real-time physiological state of the user, specifically including: the depth feature is a 24-dimensional vector, containing 24 elements reflecting the dynamic changes of the user's physiological load; the environmental interference feature is an 8-dimensional vector, containing 8 elements reflecting the environmental interference on the plateau; in order to ensure that the numerical distribution of the two types of features is consistent after splicing, a numerical consistency check needs to be performed first; the first step is to calculate the average value of the depth feature, the second step is to calculate the average value of the environmental interference feature, and the third step is to perform consistency judgment, that is, to calculate the absolute value of the difference between the average value of the depth feature and the average value of the environmental interference feature, and compare it with a preset threshold (0.15). The statistical process of the threshold is as follows: 500 groups of historical physiological data of plateau climbers at an altitude of 3000-6000 meters are selected, and the absolute value of the difference between the average value of the depth feature and the average value of the environmental interference feature in each group of data (denoted as D value) is calculated. The mean of the 500 D values is assumed to be 0.08, the standard deviation is 0.03, and the 90% quantile is 0.15. Among them, 92% of the D values are assumed to fall within the interval of 0.02-0.15, and the D values corresponding to the samples in this interval are all normal samples with matching numerical fluctuation trends of the two types of features. Therefore, 0.02-0.15 is determined as the numerical distribution dense interval, and 0.15 is taken as the threshold. After completing the numerical consistency check, the feature splicing and adjustment are performed according to the check result. If the absolute value of the difference is ≤0.15, the numerical distribution of the two types of features is consistent, and the 24 elements of the depth feature and the 8 elements of the environmental interference feature are arranged in order according to the depth feature and the environmental interference feature determined in step 101 to form a 32-dimensional feature vector. If the absolute value of the difference is >0.15, the numerical distribution of the two types of features is inconsistent, and the environmental interference feature needs to be adjusted first. The adjustment method is that each element of the environmental interference feature is added by (the average value of the depth feature - the average value of the environmental interference feature) x adjustment coefficient 0.6. The determination of the adjustment coefficient 0.6 needs to be verified by the standard deviation ratio and the distribution overlap degree. Both of them are used to judge whether the numerical distribution of the adjusted depth feature and the environmental interference feature is matched. The specific calculation and verification process is as follows: first, calculate the standard deviation of the adjusted depth feature (denoted as S1) and the standard deviation of the environmental interference feature (denoted as S2), respectively; then calculate the result by the formula standard deviation ratio = S2 ÷ S1. The closer the ratio is to 1, the more matched the dispersion degree of the two types of features is; collect the 24 elements of the adjusted depth feature and the 8 elements of the environmental interference feature, and combine them into the same set of numerical values; take the minimum value and the maximum value of the set, and evenly divide the interval from the minimum value to the maximum value into 10 equal intervals; for each interval, count the number of depth feature elements (denoted as N1) and the number of environmental interference feature elements (denoted as N2), and adjust the environmental interference feature according to the following formula: the minimum value of the interval to the maximum value of the interval is divided into 10 equal intervals, and the number of elements of the environmental interference feature in each interval is adjusted to be equal to the number of elements of the depth feature in the same interval. After adjustment, the numerical distribution of the two types of features is consistent, and the 24 elements of the depth feature and the 8 elements of the environmental interference feature are arranged in order to form a 32-dimensional feature vector. The coincidence proportion of each interval is calculated, and finally the coincidence proportions of the 10 intervals are summed up and divided by 10 to obtain the average value of the distribution coincidence degree, and the higher the average value is, the more matched the numerical value distribution ranges of the two types of features are; in the plateau environment with an altitude of 4000-5000 meters, 100 groups of samples with inconsistent feature distributions are selected to test the adjustment effect of different coefficients in the interval of 0.1-1.0, assuming that when the coefficient is 0.6, the standard deviation ratio results of the 100 groups of samples fall between 0.9-1.1, and in the distribution coincidence degree calculation, the average value of the interval coincidence proportions of the 10 groups of samples is more than 85% (for example, the average value of a certain sample is 0.865 obtained by dividing the sum of the 10 interval coincidence proportions by 10), so 0.6 is determined as the optimal adjustment coefficient; after the adjustment of the environmental interference features is completed, the features are spliced in the order of the depth features and the environmental interference features to form a 32-dimensional feature vector, which integrates the dynamic changes of the user's physiological load and the two types of key information of the plateau environmental interference, and can completely represent the real-time physiological state of the user at the current moment, such as the real-time physiological state of the user during plateau climbing and rest switching.

[0021] In this embodiment, the time correlation of multiple physiological signals is established and the time sequence features are stored through the gated recurrent unit, so that the problem of being difficult to capture the long-time sequence physiological signal coupling rule is solved, especially when the plateau climbing user switches from rest to climbing, the physiological sequence features before the state mutation can be effectively memorized, and the state judgment deviation caused by the lack of time sequence correlation is avoided; the depth features and the environmental interference features are extracted through the multi-hidden layer fully connected neural network and the nonlinear activation function, compared with the single signal processing mode, the load change information hidden in the physiological signals can be more comprehensively mined, and the plateau environmental interference factors are also integrated, so that the interference of environmental changes such as air pressure and temperature on the physiological state judgment is reduced, and the accuracy of feature extraction is improved; the real-time physiological state feature vector formed by splicing and fusing the depth features and the environmental interference features can reflect the physiological changes of the user and the environmental influence at the same time, and solves the problem of incomplete state description caused by relying on a single parameter and ignoring environmental factors.

[0022] In a preferred embodiment of the present application, the step 2 comprises: In the embodiment of the present application, in step 200, the feature vector is projected to a preset high-dimensional physiological state space, a blood oxygen saturation sub-feature vector and a respiratory rate sub-feature vector are separated and extracted from the projected feature vector, and the specific process comprises: constructing a high-dimensional physiological state space, the space is constructed based on 500 groups of historical physiological data in high-altitude mountaineering, the data covers original monitoring data of different mountaineering stages (rest, slow climbing, fast climbing, rest) at altitudes of 3000-6000 meters, including blood oxygen saturation (90%-100%), respiratory rate (12-25 times / minute), heart rate (60-150 times / minute), systolic pressure (90-160 mmHg), etc., the space is provided with 12 dimensions, the 1st dimension corresponds to the 1-minute moving average value of blood oxygen saturation, that is, the blood oxygen saturation data collected every 5 seconds within 1 minute (a total of 12 data points) are summed and divided by 12; the 2nd dimension corresponds to the fluctuation amplitude of blood oxygen saturation within 5 minutes, that is, 30 blood oxygen saturation data collected every 10 seconds within 5 minutes are used to subtract the maximum value from the minimum value; the 3rd dimension corresponds to the 10-minute change trend slope of blood oxygen saturation, that is, 20 blood oxygen saturation data collected every 30 seconds within 10 minutes are used as the sampling time (0.5, 1, 1.5……10)For the horizontal axis variable, the slope is calculated by linear regression with the blood oxygen saturation at the corresponding time point as the vertical axis variable; the 4th dimension corresponds to the 1-minute moving average of respiratory rate; the 5th dimension corresponds to the standard deviation of the interval between respiratory rate peaks; the 6th dimension corresponds to the 3-minute change rate of respiratory rate, that is, taking the 1-minute average of respiratory rate at the start of 3 minutes (denoted as F1) and the 1-minute average of respiratory rate at the end (denoted as F2), and using (F2-F1) ÷ 3; the 7th-9th dimensions correspond to the average heart rate (sum of heart rate data within 1 minute ÷ number of times collected), heart rate variability (the RR interval between consecutive heartbeats within 5 minutes of collection, that is, the time interval data between the R wave of the previous heartbeat and the R wave of the next heartbeat in the adjacent two heartbeats, and the standard deviation of these RR interval data is the heart rate variability), and the correlation between heart rate and altitude, that is, the amount of change in heart rate (the heart rate value at the end of the period minus the heart rate value at the start) ÷ the amount of change in altitude (the altitude at the end of the period minus the altitude at the start); the 10th-12th dimensions correspond to the average systolic pressure, diastolic pressure, and blood pressure fluctuation coefficient (standard deviation of systolic pressure ÷ average systolic pressure); each dimension is standardized by historical data, with a value range of 0-1, such as the 1st dimension 0 corresponding to an actual blood oxygen saturation of 90%, and 1 corresponding to 100%; the 4th dimension 0 corresponding to a respiratory rate of 12 times / minute, and 1 corresponding to 25 times / minute; the 32-dimensional feature vector obtained in step 101, including 24-dimensional physiological load features and 8-dimensional environmental interference features, is input into the above high-dimensional space, and principal component projection algorithm is used for mapping, with the specific mapping rules being that the 5th element in the 2-dimensional vector records the 10-second fluctuation value of blood oxygen saturation corresponding to the 2nd dimension of the high-dimensional space (both reflect blood oxygen saturation fluctuation-related information); the 12th element records the interval duration of 5 consecutive breaths corresponding to the 5th dimension of the high-dimensional space (both reflect respiratory interval-related information); the 8th element records the 1-minute change in heart rate corresponding to the 8th dimension of the high-dimensional space (both reflect heart rate dynamic change-related information); the 23rd element records the blood pressure change value per 100-meter altitude rise corresponding to the 11th dimension of the high-dimensional space (both reflect blood pressure and environmental correlation-related information); of the remaining 28 elements, the 24-dimensional physiological load features are matched with the 1st-3rd dimensions of the high-dimensional space reflecting blood oxygen-related information, the 4th-6th dimensions reflecting respiratory-related information, the 7th-9th dimensions reflecting heart rate-related information, and the 10th-12th dimensions reflecting blood pressure-related information, the 8-dimensional environmental interference features are matched with the 1st-3rd dimensions of the high-dimensional space affecting blood oxygen, the 4th-6th dimensions affecting respiration, the 7th-9th dimensions affecting heart rate, and the 10th-12th dimensions affecting blood pressure, and finally a 12-dimensional projected feature vector is generated, ensuring that each element is accurately mapped to the corresponding dimensional coordinate.

[0023] According to the high-dimensional space dimension identifier, the first-3 dimensional elements in the 12-dimensional projection feature vector are extracted, wherein the first-dimensional value is the 1-minute moving average value of blood oxygen saturation, such as 0.85 corresponding to actual 98%, the second-dimensional value is the fluctuation amplitude within 5 minutes, such as 0.3 corresponding to ±3% fluctuation, and the third-dimensional value is the 10-minute change trend slope, such as 0.2 corresponding to slow rise, and the 3-dimensional blood oxygen saturation sub-feature vector is arranged in the order of the first-3 dimensional elements; meanwhile, the fourth-6 dimensional elements are extracted, the fourth-dimensional value is the 1-minute moving average value of respiratory rate, such as 0.6 corresponding to 18 times / minute, the fifth-dimensional value is the peak interval standard deviation, such as 0.15 corresponding to small interval fluctuation, and the sixth-dimensional value is the 3-minute change rate, such as 0.4 corresponding to gradually accelerating, and the 3-dimensional respiratory rate sub-feature vector is arranged in the order of the fourth-6 dimensional elements.

[0024] Step 201, taking the blood oxygen saturation sub-feature vector and the respiratory rate sub-feature vector as orthogonal coordinate axes, a reference coordinate system of the high-dimensional physiological state space is established; based on the established reference coordinate system, the change trajectories of the blood oxygen saturation sub-feature vector and the respiratory rate sub-feature vector on the continuous time sequence are calculated respectively, specifically including: taking the 3-dimensional blood oxygen saturation sub-feature vector as the first coordinate axis (X-axis) of the reference coordinate system, and setting the coordinate axis scale according to the value range of the blood oxygen saturation sub-feature vector (0-1, corresponding to the normal fluctuation interval of the actual physiological significance); taking the 3-dimensional respiratory rate sub-feature vector as the second coordinate axis (Y-axis), and setting the scale according to its value range (0-1); the X-axis and the Y-axis are perpendicular to each other, and the intersection point is the origin (0, 0) of the high-dimensional physiological state space, which corresponds to the reference feature value of the human body in the resting state under the high-altitude environment, such as the sub-feature vector value of blood oxygen saturation 95% and respiratory rate 16 times / minute; taking 5 seconds as the fixed time interval, the physiological characteristic data of the user in the process of high-altitude mountaineering is collected, a total of 60 time points (covering 5 minutes of continuous monitoring) are collected, and the corresponding 3-dimensional blood oxygen saturation sub-feature vector and 3-dimensional respiratory rate sub-feature vector are obtained at each time point by the method of step 200, forming two groups of continuous time sequence data (blood oxygen sequence X1, X2, …, X60; respiratory sequence Y1, Y2, …, Y60, wherein Xn and Yn represent the sub-feature vector of the nth time point); for the blood oxygen saturation sub-feature vector, the coordinate value (taking the mean value dimension of the sub-feature vector as the coordinate representative) of each time point Xn on the X-axis is recorded in time sequence, and (1, X1 mean value), (2, X2 mean value), …, (60, X60 mean value) are connected in turn to form the continuous change trajectory of the blood oxygen saturation sub-feature vector; for the respiratory rate sub-feature vector, the coordinate value (taking the mean value dimension of the sub-feature vector) of each time point Yn on the Y-axis is also recorded, and (1, Y1 mean value), (2, Y2 mean value), …, (60, Y60 mean value) are connected in turn to form the continuous change trajectory of the respiratory rate sub-feature vector.

[0025] Step 202, according to the change trajectory, the direction difference and the change rate difference of the blood oxygen saturation change trajectory and the respiratory frequency change trajectory are calculated respectively, specifically including: the 60 time points are divided into 59 calculation units according to adjacent two time points, wherein unit 1 time point 1-2, unit 2 time point 2-3, …, unit 59 time point 59-60; for unit k (k=1 to 59), the direction vector of the blood oxygen saturation change trajectory = the blood oxygen coordinate value of the k+1th time point - the blood oxygen coordinate value of the kth time point (such as unit 1, X2 mean - X1 mean); the direction vector of the respiratory frequency change trajectory = the respiratory coordinate value of the k+1th time point - the respiratory coordinate value of the kth time point (such as unit 1, Y2 mean - Y1 mean); the angle θ between the two direction vectors in each unit is calculated by the vector angle formula After calculating cosθ, the angle θ is obtained by the inverse cosine function, such as in unit 1, vector A=0.05, vector B=0.03, θ=0°; according to the rule that the larger the angle is, the greater the direction difference is, θ is converted into a difference degree value between 0 and 1, the conversion method is difference degree value=θ÷360°×2, the difference degree values of the 59 units are summed up and then divided by 59, to obtain the overall direction difference; similarly, taking 59 adjacent time points as calculation units, for unit k, the blood oxygen change rate = (the blood oxygen coordinate value of the k+1th time point - the blood oxygen coordinate value of the kth time point) ÷ time interval (5 seconds), such as unit 1, (X2 mean - X1 mean) ÷ 5; the respiratory change rate = (the respiratory coordinate value of the k+1th time point - the respiratory coordinate value of the kth time point) ÷ 5 seconds, such as unit 1, (Y2 mean - Y1 mean) ÷ 5; ; the rate difference degree values of the 59 units are summed up and then divided by 59, to obtain the overall change rate difference.

[0026] Step 203, the direction difference and the change rate difference are weighted and fused to obtain the evolution direction difference value of the blood oxygen saturation feature vector and the respiratory frequency feature vector, specifically including: by analyzing 500 groups of physiological state early warning historical data in high altitude mountaineering, including 200 groups of physiological abnormal early warning data and 300 groups of normal physiological data, the contribution of direction difference and rate difference to physiological abnormal early warning is counted, that is, in the 200 groups of abnormal early warning data, 120 groups are triggered by the direction difference (accounting for 60%), and 80 groups are triggered by the rate difference (accounting for 40%), according to which the weight of the direction difference is determined as 0.6, the weight of the change rate difference is determined as 0.4, and the sum of the two is 1; the weighted value of the direction difference = the direction difference × 0.6; the weighted value of the change rate difference = the change rate difference × 0.4; the sum of the two weighted values is the evolution direction difference value of the blood oxygen saturation feature vector and the respiratory frequency feature vector.

[0027] In step 204, according to the evolution direction difference value, a clear boundary between the normal physiological state region, the compensatory physiological state region and the abnormal physiological state region in the high-dimensional physiological state space is divided, specifically including: Based on 500 groups of high-altitude mountaineering historical physiological data statistics, the evolution direction difference values corresponding to the normal physiological state (no obvious discomfort of human body, index coordination), the compensatory physiological state (slight adjustment of body, small amplitude difference of index) and the abnormal physiological state (obvious discomfort of body, significant difference of index) are extracted, and the first threshold value is determined to be 0.3 (the boundary between normal and compensation) and the second threshold value is determined to be 0.7 (the boundary between compensation and abnormality) by statistical distribution; when the evolution direction difference value is less than or equal to 0.3, the corresponding high-dimensional physiological state space region is the normal physiological state region, and the coordinate range of the region in the X-Y coordinate system is , corresponding to the coordination interval of blood oxygen saturation 93%-97% and respiratory rate 14-18 times / minute; when the evolution direction difference value is greater than 0.3 and less than or equal to 0.7, the corresponding region is the compensatory physiological state region, and the coordinate range is , corresponding to the adjustment interval of blood oxygen saturation 90%-93% or 97%-100% and respiratory rate 12-14 times / minute or 18-20 times / minute; when the evolution direction difference value is greater than 0.7, the corresponding region is the abnormal physiological state region, and the coordinate range is , corresponding to the abnormal interval of blood oxygen saturation <90% and respiratory rate <12 times / minute or >20 times / minute; the coordinate ranges of the three regions are ensured to have no overlap, for example, the X-axis boundary 0.6 of the normal region and the compensatory region has no intersection and no gap, for example, the compensatory region is connected with the normal region , and the boundary coordinates of each region correspond to the evolution direction difference value range one by one, realizing clear definition.

[0028] In this embodiment, by projecting the feature vector into the preset high-dimensional physiological state space and separating and extracting the key sub-feature vector, the two core physiological indexes of blood oxygen saturation and respiratory rate can be accurately focused, and irrelevant feature interference can be avoided; a reference coordinate system with the key sub-feature vector as the orthogonal coordinate axis is established, and the change trajectory on the continuous time sequence is calculated, which can intuitively present the dynamic change trend of blood oxygen saturation and respiratory rate with time, facilitate clear observation of the evolution law of the physiological indexes, and the establishment of the reference coordinate system enables the physiological data at different time points to have a unified analysis benchmark, thereby improving the effectiveness of data comparison; the direction difference degree and the change rate difference degree are calculated respectively, which can comprehensively capture the difference of the changes of blood oxygen saturation and respiratory rate from the two dimensions of direction and rate, avoid the limitation of single-dimensional analysis, and ensure that the evaluation of the change difference of the physiological indexes is more comprehensive; the evolution direction difference value is obtained by weighted fusion, which can combine the influence degree of direction and rate difference on the physiological state, so that the obtained evolution direction difference value can better reflect the real change of the physiological state; according to the evolution direction difference value, the clear boundary of different physiological state regions is divided, which can clearly define the range of different physiological states in the high-dimensional physiological state space, so that the normal, compensatory and abnormal physiological states of the human body after high altitude mountaineering have clear distinguishing standards, the physiological state type can be quickly identified, and the risk of physiological abnormality in the process of high altitude mountaineering is reduced.

[0029] In a preferred embodiment of the present application, the step 3 comprises: In the embodiment of the present application, step 300, a first state feature point is set at the center position of the normal physiological state region, a second state feature point is set on the boundary between the normal physiological state region and the compensatory physiological state region, a third state feature point is set on the boundary between the compensatory physiological state region and the abnormal physiological state region, and a fourth state feature point is set at the center position of the abnormal physiological state region, specifically comprising: determining the coordinate range of each physiological state region, which is determined based on the division result of step 204; the normal physiological state region is in a high-dimensional physiological state space with the blood oxygen saturation sub-feature vector as the X axis and the respiratory rate sub-feature vector as the Y axis, the X axis coordinate range is 0.6 to 0.9, and the Y axis coordinate range is 0.5 to 0.8; the compensatory physiological state region X axis coordinate range is 0.4 to 0.6 and 0.9 to 1.0, and the Y axis coordinate range is 0.3 to 0.5 and 0.8 to 1.0; the abnormal physiological state region X axis coordinate range is 0 to 0.4, and the Y axis coordinate range is 0 to 0.3 and 1.0 to 1.2; calculating the center coordinates of the normal physiological state region, the X axis center coordinate is the sum of the maximum and minimum values of the normal region X axis divided by 2, and the Y axis center coordinate is the sum of the maximum and minimum values of the normal region Y axis divided by 2, to obtain the center coordinates assumed as (0.75, 0.65), which is set as the first state feature point; then setting the second state feature point, selecting the midpoint of the boundary between the normal physiological state region and the compensatory physiological state region as the point, selecting the adjacent boundary of the two regions on the X axis, i.e. the minimum value 0.6 of the normal region X axis and the maximum value 0.6 of the compensatory region X axis, and the adjacent boundary on the Y axis, i.e. the minimum value 0.5 of the normal region Y axis and the maximum value 0.5 of the compensatory region Y axis, taking the midpoint coordinate, i.e. the X axis coordinate is (0.6+0.6)÷2; the Y axis coordinate is (0.5+0.5)÷2, to obtain the midpoint coordinate (0.6, 0.5) set as the second state feature point; then setting the third state feature point, selecting the midpoint of the boundary between the compensatory physiological state region and the abnormal physiological state region as the point, selecting the adjacent boundary of the two regions on the X axis, i.e. the minimum value 0.4 of the compensatory region X axis and the maximum value 0.4 of the abnormal region X axis, and the adjacent boundary on the Y axis, i.e. the minimum value 0.3 of the compensatory region Y axis and the maximum value 0.3 of the abnormal region Y axis, taking the midpoint coordinate, i.e. the X axis coordinate is (0.4+0.4)÷2; the Y axis coordinate is (0.3+0.3)÷2, to obtain the midpoint coordinate (0.4, 0.3) set as the third state feature point; finally, setting the fourth state feature point, calculating the center coordinates of the abnormal physiological state region, the X axis center coordinate is the sum of the maximum and minimum values of the abnormal region X axis divided by 2; the Y axis center coordinate is the sum of the maximum and minimum values of the abnormal region Y axis divided by 2, to obtain the center coordinates assumed as (0.2, 0.15), which is set as the fourth state feature point.

[0030] Step 301, calculate the spatial distribution characteristics of the four state feature points in the high-dimensional physiological state space with the blood oxygen saturation sub-feature vector and the respiratory rate sub-feature vector as the coordinate axes, including the relative distance and position relationship between the state feature points, specifically including: based on the four state feature point coordinates determined in step 300, the first state feature point coordinate is marked as P1(0.75, 0.65), the second is marked as P2(0.6, 0.5), the third is marked as P3(0.4, 0.3), and the fourth is marked as P4(0.2, 0.15), calculate the relative distance and position relationship between each point; when calculating the relative distance between each state feature point, the Euclidean distance calculation method is used to calculate the relative distance between P1 and P2; the relative distances between P2 and P3, P3 and P4, P1 and P3, P2 and P4, and P1 and P4 are calculated in the same way; when analyzing the position relationship of each state feature point, observe the coordinate change trend of each point on the X-axis and Y-axis, P1 to P2, then to P3, and finally to P4, assuming that the X-axis coordinate gradually decreases from 0.75 to 0.6, 0.4, and finally to 0.2, showing a gradually decreasing trend; the Y-axis coordinate gradually decreases from 0.65 to 0.5, 0.3, and finally to 0.15, also showing a gradually decreasing trend, and the arrangement of each point on the coordinate axis shows a linear distribution, transitioning from the normal physiological state region to the compensatory and abnormal physiological state region in turn.

[0031] Step 302, based on the spatial distribution characteristics of the state feature points, a correction factor representing the current physiological state spatial distribution characteristics is generated; the feature vector is weighted and corrected using the correction factor to obtain a corrected feature vector, which specifically includes: based on the relative distances of the state feature points calculated in step 301, the relative distance between P1 and P4 is selected as the reference distance, and the relative distance between P2 and P3 is selected as the reference distance, and the value of the correction factor is calculated, that is, the reference distance and the reference distance are added to obtain the sum of the reference distance and the reference distance; then divide the sum by 2 to obtain the average distance; finally, the average distance x 0.8 (the coefficient is determined by analyzing 500 groups of historical physiological data on the plateau, specifically, for each group of 500 groups of data, different coefficients in the interval of 0.1-1.0 are used to generate correction factors, and the physiological state judgment results corresponding to the corrected feature vectors are compared with the actual physiological state monitored at the same period, such as normal, compensation, and abnormal clinical judgment results. The number of samples with consistent judgment results and actual state in each group of data is counted, and the total number of samples in the group is divided to obtain the single-group matching degree, and the average matching degree of 500 groups of data is calculated. Assuming that the coefficient is 0.8, the average matching degree of 500 groups of data is the highest, which can match the change rule of physiological state in plateau environment), the result is the correction factor; then determine the feature vector to be corrected, which is the 32-dimensional feature vector used when dividing the physiological state region in step 204, containing 24-dimensional physiological load feature elements and 8-dimensional environmental interference feature elements; multiply the correction factor by each element in the 32-dimensional feature vector to obtain the corrected value of each element. For the first element to the 24th element (physiological load feature elements) in the feature vector, the corrected value of each element is equal to the original value of the element x the correction factor; for the 25th element to the 32nd element (environmental interference feature elements), the corrected value of each element is also equal to the original value of the element x the correction factor; finally, arrange the corrected values of all elements in the order of the original 32-dimensional feature vector, that is, arrange the first corrected element to the 24th corrected element in turn, and then arrange the 25th corrected element to the 32nd corrected element, to form a new vector, which is the corrected feature vector.

[0032] The embodiment can accurately mark the core positions and transition positions of each physiological state by setting feature points at different physiological state regions and region boundaries, and avoid calculation deviation caused by ambiguous reference points; the relative distance and position relationship of each feature point can clearly present the distribution rule of different physiological state regions in high-dimensional space, and clearly show the spatial correlation between normal, compensatory and abnormal physiological states, so that the correction factor can accurately reflect the spatial distribution characteristics of the physiological state; the correction factor is generated based on the spatial distribution characteristics, and the feature vector is weighted and corrected, which can eliminate the influence of environmental interference and physiological index fluctuation in the plateau environment on the feature vector, and make the corrected feature vector more consistent with the real physiological state of the user.

[0033] In a preferred embodiment of the application, the step 4 comprises: In the embodiment of the application, in step 400, the corrected feature vector is calculated to generate a control instruction set comprising a target injection frequency, a driving pressure and a time ratio of inspiration and expiration, specifically comprising: determining a mapping relationship between the corrected feature vector and the control parameters, the mapping relationship being established based on the matching results of 500 groups of historical physiological data and oxygen supply parameters in plateau climbing, and the specific establishment process being that for each group of the 500 groups of historical data, the corresponding physiological feature vector is extracted, which is consistent with the 32-dimensional feature vector structure and the optimal oxygen supply parameters, i.e. the target injection frequency, the driving pressure and the time ratio of inspiration and expiration determined by the actual physiological feedback of the climber and the standard condition of blood oxygen and respiratory index; the correlation analysis method is used to calculate the correlation coefficient of each element in the 32-dimensional feature vector and each oxygen supply parameter, taking the correlation coefficient of a certain dimensional element and the target injection frequency as an example, the average value of the 500 elements of the dimensional element is calculated and recorded as Xmean, and the average value of the 500 corresponding values of the target injection frequency is calculated and recorded as Ymean; for each group of data, (the element value of the group-Xmean) x (the target injection frequency of the group-Ymean) is calculated, and the sum of the 500 groups of results is obtained as the numerator; then the square sum of (the element value of the group-Xmean) and the square sum of (the target injection frequency of the group-Ymean) are calculated, and the square root of the product of the two square sums is obtained as the denominator; finally, the numerator is divided by the denominator, and the result is the correlation coefficient of the dimensional element and the target injection frequency; the correlation coefficients of each dimensional element and the driving pressure and the time ratio of inspiration and expiration are calculated in the same way, assuming that the average correlation coefficient of the first 8-dimensional elements and the target injection frequency is 0.82 (much higher than other dimensions), the average correlation coefficient of the 9th-16th-dimensional elements and the driving pressure is 0.78, and the average correlation coefficient of the 17th-24th-dimensional elements and the time ratio of inspiration and expiration is 0.75, according to which the first 8-dimensional elements (physiological load core features) correspond to the target injection frequency, the 9th-16th-dimensional elements (environmental interference adaptation features) correspond to the driving pressure, and the 17th-24th-dimensional elements (physiological state coordination features) correspond to the time ratio of inspiration and expiration.

[0034] The first to eighth dimension elements of the corrected feature vector are extracted, the average of the eight elements is calculated, the average is multiplied by the frequency coefficient 15 to obtain the target injection frequency; wherein the frequency coefficient 15 is determined by historical data statistics, that is, the average of the first to eighth dimension elements and the corresponding optimal target injection frequency are extracted from 500 groups of historical data, a corresponding relationship table of the average and the target injection frequency is established, it is assumed that the optimal target injection frequency increases by 1.5 times per minute for every increase of 0.1 in the average, for example, the average of 0.2 corresponds to 3 times per minute, the average of 0.3 corresponds to 4.5 times per minute, and the average of 0.6 corresponds to 9 times per minute, and the frequency coefficient is calculated by linear fitting as 1.5 ÷ 0.1 = 15; the ninth to sixteenth dimension elements of the corrected feature vector are extracted, the weighted sum of the eight elements is calculated, wherein the ninth to twelfth dimension elements have a weight of 0.6 (environmental interference core influence), and the thirteenth to sixteenth dimension elements have a weight of 0.4 (environmental interference secondary influence), and the weighted sum is multiplied by the pressure coefficient 20 to obtain the driving pressure; wherein the pressure coefficient 20 is determined by historical data statistics, specifically, the weighted sum of the ninth to sixteenth dimension elements and the corresponding optimal driving pressure are extracted from 500 groups of historical data, it is assumed that the optimal driving pressure is 6 kPa when the weighted sum is 0.3, 8 kPa when the weighted sum is 0.4, and 14 kPa when the weighted sum is 0.7, and the pressure coefficient is calculated by linear fitting as 6 ÷ 0.3 = 20; the seventeenth to twenty-fourth dimension elements of the corrected feature vector are extracted, the difference between the maximum and the minimum of the eight elements is calculated, the difference is added to the reference ratio 1.2 (highland normal inhalation-exhalation phase time ratio to the reference), to obtain the inhalation-exhalation phase time ratio; the target injection frequency, the driving pressure and the inhalation-exhalation phase time ratio obtained by solving are arranged in the order of injection frequency, driving pressure and inhalation-exhalation phase time ratio to form a control instruction set.

[0035] Step 401, according to the control instruction set drive piezoelectric ceramic high frequency valve to execute corresponding parameters of intermittent oxygen injection, and through the micro flow sensor and pressure sensor real-time collection actual output gas flow data and pressure data, specific including: the target injection frequency in the control instruction set determines the on-off cycle of the high frequency valve, cycle time = 60 seconds ÷ target injection frequency, assuming that the cycle time is 10.91 seconds, that is, the high frequency valve completes one "open to close" action every 10.91 seconds; the driving pressure determines the opening amplitude of the high frequency valve, the greater the driving pressure, the greater the displacement of the high frequency valve core, the greater the oxygen injection channel opening; the inspiration-expiration time ratio determines the time allocation of the inspiration period and the expiration period in a single injection, for example, the inspiration-expiration time ratio is 1.44:1, and the total time length of a single injection is 2 seconds (according to the plateau breathing rhythm), then the inspiration period time length is (1.44 ÷ (1.44 + 1)) × 2, and the expiration period time length is 2 - inspiration period time length; according to the above driving logic, send an electrical signal to the piezoelectric ceramic high frequency valve to control the high frequency valve to open at the set opening amplitude during the inspiration period, output oxygen, and close during the expiration period to realize intermittent oxygen injection; the micro flow sensor is installed on the high frequency valve oxygen output pipeline, and the instantaneous flow data of oxygen in each injection process is collected at an interval of 0.1 seconds; the pressure sensor is installed at the inlet of the high frequency valve, and the instantaneous pressure data in each injection process is collected at an interval of 0.1 seconds; the collected flow data and pressure data in each injection cycle are stored respectively to form a group of actual output data.

[0036] Step 402, compare the actual gas flow data, pressure data with the corresponding target values in the target control instruction set, calculate the gas flow error signal and the pressure error signal, specific including: the target flow value is determined based on the target injection frequency and the driving pressure in the control instruction set, that is, target flow value = driving pressure × 0.1 (pressure flow conversion coefficient) + target injection frequency × 0.05 (frequency flow compensation coefficient); take the average value of the actual flow data collected in a certain injection cycle to get the actual average flow; actual average flow - target flow value = flow error; flow error × error amplification coefficient 2 (to ensure that error 0.1 L / min corresponds to error signal 0.2), get the gas flow error signal; take the average value of the actual pressure data collected in a certain injection cycle to get the actual average pressure; actual average pressure - target pressure value = pressure error; pressure error × error amplification coefficient 1.5 (to ensure that error 0.1 kPa corresponds to error signal 0.15), get the pressure error signal.

[0037] Step 403, feeds back the gas flow error signal and the pressure error signal to the adaptive neural network, adjusts the connection weights of the neural network by calculating the partial derivatives of the error signal with respect to the connection weights of the neural network, and adjusts the internal state parameters of the gated cyclic unit according to the time series characteristics of the error signal, including updating the gate weights and resetting the gate weights. Specifically, the gas flow error signal and the pressure error signal calculated in step 402 are fed back to the adaptive neural network. The neural network has the same structure as that in step 1. The network has a three-layer structure, the input layer is two nodes of the gas flow error signal and the pressure error signal obtained in step 402, the hidden layer is eight nodes, including two gated cyclic units, and the output layer is one node of the correction factor adjustment amount; the initial value of the connection weight between each layer is determined based on 500 sets of historical error data training, that is, a single set of data is extracted from the 500 sets of historical error data. For example, if the gas flow error signal is -2.54 and the pressure error signal is -0.27, the actual value of the output layer correction factor adjustment is 0.08. First, set the initial random weights from the input layer to the hidden layer. For example, the corresponding weights of the first hidden layer node are 0.4 and 0.3, and the corresponding weights of the second hidden layer node are 0.5 and 0.4. The random weights of the remaining 6 hidden layer nodes are set in the same way. The input value of the first hidden layer node = gas flow error signal -2.54 × corresponding weight 0.4 + pressure error signal -0.27 × corresponding weight 0.3. The input values ​​of the remaining 7 hidden layer nodes are calculated in the same way. Substitute the hidden layer node input value into the Sigmoid activation function to obtain the hidden layer node output value. For example, the output value of the first node is 0.25. The output layer prediction value = the output value of each hidden layer node × the corresponding initial random weight from the hidden layer to the output layer (such as 0.5) and then sum them. The weights are adjusted by the gradient descent method. For example, the gas flow error signal from the input layer to the first hidden layer corresponds to the weight, and the adjustment amount = error × learning rate 0.01 × gas flow error signal × derivative of the hidden layer node activation function. The new weight = original weight 0.4 + adjustment amount. The weight adjustment process is iterated repeatedly until the error is stably minimized. Finally, the initial values ​​of the connection weights from the input layer to the hidden layer are assumed to be 0.5, 0.4, 0.6, 0.3, 0.55, 0.45, 0.58, and 0.4.

[0038] The activation function of the hidden layer node adopts a Sigmoid function consistent with step 100; the partial derivative of the gas flow error signal to each connection weight from the input layer to the hidden layer is calculated, and the partial derivative = gas flow error signal * activation value of the corresponding hidden layer node; the partial derivative of the pressure error signal to each connection weight from the input layer to the hidden layer is calculated, and the calculation method is consistent with the flow error signal; the two partial derivatives are added to obtain the total partial derivative; the original connection weight from the input layer to the hidden layer is adjusted to the adjusted connection weight = original connection weight - total partial derivative * learning rate 0.01; the gas flow error signal and the pressure error signal in 10 consecutive injection cycles are collected to form an error time sequence; the sliding average of the sequence is calculated, and the calculation method is to take an average every 3 cycles; if the absolute value of the sliding average is greater than 0.5, it is determined that the error trend is obvious, at this time, the update gate weight is adjusted, the initial weight of the update gate is 0.3, and the adjustment step of the update gate weight is added when adjusting, 0.02, which is used to moderately increase the update gate weight when the error trend is obvious, and the adjusted weight is 0.3+0.02=0.32; if the absolute value of the sliding average is less than 0.2, it is determined that the error is stable, at this time, the reset gate weight is adjusted, the initial weight of the reset gate is 0.4, and the adjustment step of the reset gate weight is subtracted when adjusting, 0.01, which is used to moderately reduce the reset gate weight when the error is stable, and the adjusted weight is 0.4-0.01=0.39.

[0039] In this embodiment, the oxygen supply control instruction set is calculated by correcting the feature vector, which can accurately match the real-time physiological state of the user with the oxygen supply parameters, avoid the problems of insufficient or excessive oxygen supply in the plateau environment, and ensure that the oxygen injection frequency and pressure are matched with the breathing demand of the mountaineer; the high-frequency valve is driven to execute injection and real-time data acquisition, which can monitor the actual output state of the oxygen supply system in real time, and capture the dynamic changes of the flow and pressure in time; the flow and pressure error signals are calculated, which can accurately quantify the deviation between the actual oxygen supply and the target demand, determine the system adjustment direction, and avoid the decline of the oxygen supply effect caused by the deviation accumulation; the connection weight and the gate parameter are adjusted through the neural network, which can dynamically optimize the system response according to the error change.

[0040] In a preferred embodiment of the present application, the step 5 comprises: In the embodiment of the present application, in step 500, when the user motion intensity mutation is detected through the three-axis motion acceleration data, the blood oxygen saturation value at each time point in the next 30 seconds is predicted based on the blood oxygen saturation history data, the respiratory rate history data and the electrocardiogram signal history data memorized by the gated recurrent unit, through the time series prediction function of the gated recurrent unit, to form a blood oxygen saturation value sequence, specifically including: determining the detection standard of the motion intensity mutation, the three-axis motion acceleration data (X-axis, Y-axis, Z-axis) is collected in real time through the wearable device, the collection interval is 0.5 seconds, the acceleration vector sum change amount in the continuous two collection intervals is calculated, the sum of the square of the X-axis acceleration, the square of the Y-axis acceleration and the square of the Z-axis acceleration in the previous interval is calculated first, and the square root is taken to obtain the acceleration vector sum of the previous interval; the acceleration vector sum of the next interval is calculated in the same way; the acceleration vector sum change amount is obtained by subtracting the acceleration vector sum of the previous interval from the acceleration vector sum of the next interval; if the change amount is greater than the preset threshold value 0.8 m / s² (the intensity mutation critical value based on 500 groups of motion data statistics of high altitude mountaineering, the specific statistical process is to select 180 groups of mutation data from 500 groups of data, in which the user motion intensity changes from low intensity such as slow walking and short stay to medium-high intensity such as rapid climbing, heavy load walking, or from medium-high intensity to high intensity such as steep slope sprint climbing, calculate the acceleration vector sum change amount at each mutation, and take the average value of the 180 change amounts as 0.72 m / s², then 0.72 m / s²×safety factor 1.1 (adapted to the risk redundancy demand of high altitude), 0.8 m / s² is obtained as the intensity mutation critical value), it is determined that the user motion intensity mutation occurs, such as from slow climbing to rapid climbing, from static to climbing, etc.; then the historical data memorized by the gated recurrent unit is called, the gated recurrent unit is the gated recurrent unit whose parameters are adjusted in step 403, and the historical data memorized by the gated recurrent unit includes the blood oxygen saturation history data in the past 5 minutes, 1 second collection, a total of 300 data points, the respiratory rate history data, 2 second collection, a total of 150 data points, the electrocardiogram signal history data, 0.5 second collection, a total of 600 data points, and the historical data is extracted in chronological order to form an ordered historical data set.

[0041] The future 30 seconds of blood oxygen saturation is calculated by the time series prediction function of the gated recurrent unit, the prediction interval is 1 second, a total of 30 time points of blood oxygen saturation values need to be predicted. When predicting, first input the blood oxygen saturation historical data, respiratory rate historical data, and electrocardiogram signal historical data in the historical data set into the gated recurrent unit according to the time step. The gated recurrent unit selectively remembers and forgets the historical data by updating the gate weight and resetting the gate weight (the adjusted weight in step 403, the updated gate weight is 0.32, and the reset gate weight is 0.39). That is, the historical data with obvious change trend in the recent 1 minute, such as continuous decrease of blood oxygen and acceleration of respiratory rate, is enhanced memory, and the early stable data is appropriately forgotten. Then, based on the key historical data in memory, the correlation degree of each prediction time point with historical data is calculated by circular calculation, that is, the interval reciprocal of the prediction time point with the historical data time point is multiplied by the data change trend similarity (the prediction time point corresponding to the future 5 seconds of prediction segment and the historical 5 seconds of data segment to be matched are divided into 5 small segments according to 1 second, the change direction of data in each small segment is calculated, that is, the data value of the last 1 second is subtracted from the data value of the first 1 second, if the result is greater than 0, it is recorded as 1 (up), if the result is less than 0, it is recorded as -1 (down), if the result is equal to 0, it is recorded as 0 (stable), the number of small segments with consistent change direction in the two data segments is counted, and the similarity is obtained by dividing the consistent number by the total number of small segments 5, the similarity range is 0 to 1). Finally, the blood oxygen saturation value of each prediction time point is calculated according to the correlation degree, for example, when predicting the blood oxygen value of the first second, the historical blood oxygen data in the past 10 seconds is taken, the correlation degree is sorted from high to low, and the average value of the first 5 high correlation degree data is taken as the prediction value of the first second. The blood oxygen saturation values of each 1 second time point in the future 30 seconds are calculated in the same way, and 30 prediction values are arranged in time sequence to form a blood oxygen saturation value sequence.

[0042] Step 501, according to the blood oxygen saturation value sequence, adjust the generation strategy of the control instruction set, that is, when the blood oxygen saturation value sequence shows a downward trend, increase the target injection frequency and driving pressure by a preset proportion, when the blood oxygen saturation value sequence shows an upward trend, decrease the target injection frequency and driving pressure by a preset proportion; at the same time, according to the change trend of the respiratory frequency historical data memorized by the gating cycle unit, the inspiration-expiration phase time ratio is adjusted correspondingly to realize the matching of the oxygen output parameters and the real-time physiological needs of the user, which specifically includes: judging the change trend of the blood oxygen saturation value sequence, that is, calculating the value difference of the adjacent two time points in the blood oxygen saturation value sequence, a total of 29 value differences; count the number of negative numbers in the difference value, if the proportion of the number of negative numbers in the total number of difference values is greater than 60%, it is determined that the sequence shows a downward trend; if the proportion of the number of positive numbers in the total number of difference values is greater than 60%, it is determined that the sequence shows an upward trend; if the proportion of positive and negative numbers is less than 60%, it is determined that the sequence shows a stable trend; then adjust the target injection frequency and driving pressure, the preset increase proportion is 15%, the decrease proportion is 10%, when it is determined that the sequence shows a downward trend, the new target injection frequency = the original target injection frequency (the frequency calculated in step 400) + the original target injection frequency x 15%; the new driving pressure = the original driving pressure (the pressure calculated in step 400) + the original driving pressure x 15%; when it is determined that the sequence shows an upward trend, the new target injection frequency = the original target injection frequency - the original target injection frequency x 10%; the new driving pressure = the original driving pressure - the original driving pressure x 10%; if the sequence shows a stable trend, the original target injection frequency and driving pressure remain unchanged; then adjust the inspiration-expiration phase time ratio, first extract the respiratory frequency historical data memorized by the gating cycle unit in the past 2 minutes, a total of 60 data points, calculate the average value of adjacent 10 data points, get 6 segmented average values; count the change trend of the segmented average values, if the number of times that the average value of the next segment is greater than the average value of the previous segment by 0.5 times / minute or more is greater than 3 times, it is determined that the respiratory frequency shows an accelerating trend; if the number of times that the average value of the next segment is less than the average value of the previous segment by 0.5 times / minute or more is greater than 3 times, it is determined that the respiratory frequency shows a slowing trend; when the respiratory frequency shows an accelerating trend, the new inspiration-expiration phase time ratio = the original inspiration-expiration phase time ratio (the ratio calculated in step 400) + 15%; when the respiratory frequency shows a slowing trend, the new inspiration-expiration phase time ratio = the original inspiration-expiration phase time ratio - 10%; if the respiratory frequency shows a stable trend, the original inspiration-expiration phase time ratio remains unchanged; finally, arrange the adjusted new target injection frequency, new driving pressure and new inspiration-expiration phase time ratio in the order of injection frequency, driving pressure and inspiration-expiration phase time ratio to form the adjusted control instruction set.

[0043] The embodiment can perceive the blood oxygen fluctuation of the highland mountaineer caused by the change of the exercise intensity in advance, and avoid the lag problem of adjusting the oxygen supply after the actual blood oxygen decreases; the injection frequency and the driving pressure are adjusted according to the blood oxygen trend, and the inhalation-exhalation phase time ratio is adjusted according to the respiratory frequency trend, so that the oxygen supply parameters can dynamically match the physiological needs after the exercise intensity changes, the situation that the oxygen supply is insufficient when the exercise intensity increases or the oxygen supply is excessive when the exercise intensity decreases can be avoided, and the real-time and accuracy of the oxygen supply in the highland environment are improved; the whole process relies on the historical data memorized by the gated recurrent unit, ensures that the prediction and adjustment are based on the real physiological change law of the mountaineer, avoids the blind adjustment that deviates from the individual physiological characteristics, makes the oxygen supply control more suitable for the individual differences of the highland mountaineer, and guarantees the oxygen supply adaptation effect in different exercise states.

[0044] As shown in Figure 2 The embodiment of the present application also provides an intermittent injection ventilation nasal oxygen catheter intelligent control system, which comprises: An extraction module is used for inputting the multi-modal sensor data into an adaptive neural network of a gated recurrent unit to perform feature extraction and state memory, so as to obtain a feature vector; A projection module is used for mapping the feature vector to a high-dimensional physiological state space, determining a state reference coordinate based on the blood oxygen saturation and the respiratory frequency feature vector, calculating the evolution direction difference of the blood oxygen saturation feature vector and the respiratory frequency feature vector, and dividing a state boundary; A correction module is used for setting four state feature points in the state boundary, generating a correction factor based on the spatial distribution characteristics of the state feature points, correcting the feature vector by using the correction factor, and obtaining a corrected feature vector; A feedback module is used for calculating the corrected feature vector to generate a control instruction set containing a target injection frequency, a driving pressure and an inhalation-exhalation phase time ratio; driving a piezoelectric ceramic high-frequency valve to perform intermittent oxygen injection according to the control instruction set, and obtaining actual gas flow and pressure data through a sensor; comparing the actual gas flow and pressure data with the target control instruction set to generate an error signal feedback to the neural network to adjust the network parameters; A mutation module is used for predicting the blood oxygen saturation change trend in a future preset time according to the physiological state historical data memorized by the gated recurrent unit when detecting that the user's exercise intensity changes, and adjusting the generation strategy of the control instruction set to realize the matching of the oxygen output and the real-time physiological needs.

[0045] It should be noted that the system corresponds to the above method, and all implementation manners in the above method embodiment are applicable to this embodiment, and the same technical effects can be achieved.

[0046] The embodiment of the present application also provides a computing device, comprising a processor, a memory storing a computer program, the computer program being executed by the processor to perform the method as described above. All implementation manners in the above method embodiment are suitable for this embodiment and can achieve the same technical effects.

[0047] The embodiment of the present application also provides a computer readable storage medium storing instructions, which, when executed on a computer, cause the computer to perform the method as described above. All implementation manners in the above method embodiment are suitable for this embodiment and can achieve the same technical effects.

[0048] The above is the preferred embodiment of the present application, it should be pointed out that, for those skilled in the art, without departing from the principles of the present application, can make a number of improvements and refinements, these improvements and refinements should also be considered as the protection scope of the present application.

Claims

1. An intelligent control method for a nasal oxygen inhalation catheter capable of intermittent jet ventilation, characterized in that: The method comprises: Step 1: Input the multimodal sensor data into the adaptive neural network embedded in the gated recurrent unit for feature extraction and state memory to obtain the feature vector; Step 2: Map the feature vector to a high-dimensional physiological state space and determine the state reference coordinate based on the blood oxygen saturation and respiratory rate feature vectors; calculate the evolution direction difference between the blood oxygen saturation feature vector and the respiratory rate feature vector and divide the state boundary; Step 3: Set four state feature points within the state boundary, generate correction factors based on the spatial distribution characteristics of the state feature points, and use the correction factors to correct the feature vector to obtain a corrected feature vector; Step 4: The corrected eigenvector is solved to generate a control instruction set containing the target injection frequency, driving pressure, and inhalation-exhalation time ratio. According to the control instruction set, the piezoelectric ceramic high-frequency valve is driven to perform intermittent oxygen injection. Actual gas flow and pressure data are obtained through sensors. The actual gas flow and pressure data are compared with the target control instruction set to generate an error signal that is fed back to the neural network to adjust the network parameters. Step 5: When a sudden change in the user's exercise intensity is detected, the blood oxygen saturation change trend within a preset time period is predicted based on the physiological state historical data memorized by the gated recurrent unit, and the generation strategy of the control instruction set is adjusted to achieve a match between oxygen output and real-time physiological needs.

2. The intelligent control method for the intermittent jet ventilation nasal oxygen inhalation catheter according to claim 1 is characterized in that: The process of acquiring multimodal sensor data is as follows: Multimodal sensors are used to collect user physiological signals and plateau environmental parameters in real time. The physiological signals include blood oxygen saturation, respiratory waveform, electrocardiogram signal and three-axis motion acceleration. The environmental parameters include atmospheric pressure, ambient temperature and airflow velocity. The user physiological signals and the plateau environmental parameters are synchronized and time-aligned to obtain multimodal sensor data.

3. The intelligent control method for the intermittent jet ventilation nasal oxygen inhalation catheter according to claim 2, characterized in that: The step 1 comprises: The multimodal sensor data is input into an adaptive neural network embedded with a gated recurrent unit in a time series. The gated recurrent unit inside the adaptive neural network establishes the temporal correlation between the blood oxygen saturation data, respiratory waveform data, electrocardiogram data, and triaxial motion acceleration data, and stores the temporal variation characteristics of the physiological signals. The time series variation features are input into a fully connected neural network containing multiple hidden layers in the adaptive neural network, and transformed through the nonlinear activation functions of the multiple hidden layers to extract deep features reflecting the dynamic changes of physiological load and environmental interference features; The depth features and environmental interference features are fused through feature splicing operations to obtain a feature vector representing the user's real-time physiological state.

4. The intelligent control method for the intermittent jet ventilation nasal oxygen inhalation catheter according to claim 3 is characterized in that: The step 2 includes: Projecting the feature vector into a preset high-dimensional physiological state space, separating and extracting the blood oxygen saturation sub-feature vector and the respiratory rate sub-feature vector from the projected feature vector; Using the blood oxygen saturation sub-eigenvector and the respiratory rate sub-eigenvector as orthogonal coordinate axes, a reference coordinate system of the high-dimensional physiological state space is established. Based on the established reference coordinate system, the change trajectories of the blood oxygen saturation sub-eigenvector and the respiratory rate sub-eigenvector in the continuous time series are calculated respectively. According to the change trajectory, the direction difference and change rate difference between the blood oxygen saturation change trajectory and the respiratory rate change trajectory are calculated respectively; The direction difference and the change rate difference are weightedly fused to obtain the evolution direction difference value of the blood oxygen saturation feature vector and the respiratory rate feature vector; According to the difference value of the evolution direction, clear boundaries between the normal physiological state region, the compensatory physiological state region and the abnormal physiological state region are divided in the high-dimensional physiological state space.

5. The intelligent control method for the intermittent jet ventilation nasal oxygen inhalation catheter according to claim 4 is characterized in that: The step 3 includes: A first state characteristic point is set at the center of the normal physiological state area, a second state characteristic point is set at the boundary between the normal physiological state area and the compensatory physiological state area, a third state characteristic point is set at the boundary between the compensatory physiological state area and the abnormal physiological state area, and a fourth state characteristic point is set at the center of the abnormal physiological state area. Calculate the spatial distribution characteristics of the four state feature points in the high-dimensional physiological state space with the blood oxygen saturation sub-feature vector and the respiratory rate sub-feature vector as coordinate axes, including the relative distance and position relationship between the state feature points; Based on the spatial distribution characteristics of the state feature points, a correction factor representing the spatial distribution characteristics of the current physiological state is generated; the correction factor is used to perform weighted correction on the feature vector to obtain a corrected feature vector.

6. The intelligent control method for the intermittent jet ventilation nasal oxygen inhalation catheter according to claim 5, characterized in that: The step 4 comprises: The corrected eigenvector is solved to generate a control instruction set including target injection frequency, driving pressure and inspiration-expiration phase time ratio; According to the control instruction set, the piezoelectric ceramic high-frequency valve is driven to perform intermittent oxygen injection with corresponding parameters, and the actual output gas flow and pressure data are collected in real time through the micro flow sensor and pressure sensor; Compare the actual gas flow rate data and pressure data with the corresponding target values ​​in the target control instruction set, and calculate the gas flow rate error signal and pressure error signal; The gas flow error signal and pressure error signal are fed back to the adaptive neural network. The connection weights of the neural network are adjusted by calculating the partial derivatives of the error signal with respect to the connection weights of the neural network. The internal state parameters of the gated recurrent unit are adjusted according to the time series characteristics of the error signal, including updating the gate weights and resetting the gate weights.

7. The intelligent control method for the intermittent jet ventilation nasal oxygen inhalation catheter according to claim 6, characterized in that: The step 5 comprises: When a sudden change in the user's exercise intensity is detected through the three-axis motion acceleration data, the blood oxygen saturation value at each time point in the next 30 seconds is predicted through the gated recurrent unit's time series prediction function based on the historical blood oxygen saturation data, respiratory rate data, and ECG signal data stored in the gated recurrent unit, forming a blood oxygen saturation value sequence; According to the blood oxygen saturation value sequence, the generation strategy of the control instruction set is adjusted. That is, when the blood oxygen saturation value sequence shows a downward trend, the target injection frequency and driving pressure are increased according to the preset ratio; when the blood oxygen saturation value sequence shows an upward trend, the target injection frequency and driving pressure are reduced according to the preset ratio; at the same time, according to the changing trend of the historical respiratory rate data memorized by the gated cycle unit, the respiratory and expiratory phase time ratio is adjusted accordingly to achieve the matching of oxygen output parameters with the user's real-time physiological needs.

8. An intelligent control system for a nasal oxygen inhalation catheter capable of intermittent jet ventilation, the system implementing the method according to any one of claims 1 to 7, characterized in that: include: The extraction module is used to input multimodal sensor data into the adaptive neural network embedded in the gated recurrent unit for feature extraction and state memory to obtain a feature vector; A projection module is used to map the feature vector to a high-dimensional physiological state space, determine the state reference coordinates based on the blood oxygen saturation and respiratory rate feature vectors, calculate the evolution direction difference between the blood oxygen saturation feature vector and the respiratory rate feature vector, and divide the state boundary; A correction module is used to set four state feature points within the state boundary, generate correction factors based on the spatial distribution characteristics of the state feature points, and use the correction factors to correct the feature vector to obtain a corrected feature vector; A feedback module is used to solve the corrected eigenvector to generate a control instruction set containing the target injection frequency, driving pressure, and inhalation-exhalation time ratio; according to the control instruction set, the piezoelectric ceramic high-frequency valve is driven to perform intermittent oxygen injection, and actual gas flow and pressure data are obtained through sensors. The actual gas flow and pressure data are compared with the target control instruction set, and an error signal is generated and fed back to the neural network to adjust the network parameters; The mutation module is used to predict the blood oxygen saturation change trend within a preset time in the future based on the physiological state historical data memorized by the gated recurrent unit when a sudden change in the user's exercise intensity is detected, and adjust the generation strategy of the control instruction set to achieve matching of oxygen output with real-time physiological needs.

9. A computing device, characterized in that include: one or more processors; A storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a program, which, when executed by a processor, implements the method according to any one of claims 1 to 7.

Images