Intelligent dynamic monitoring device and method for chronic obstructive pulmonary disease based on wireless sensing

Abstract

The invention belongs to the technical field of medical equipment, and particularly provides an intelligent dynamic monitoring device and method for chronic obstructive pulmonary disease based on wireless sensing, which get rid of dependence on an expiration examination mode, reduce the difficulty of examination operation, and reflect the severity and variability of airway obstruction in other signal monitoring modes. Comprehensive evaluation is carried out in combination with extrapulmonary indexes (such as lung volume change, heart rate, blood oxygen, breathing sound and gait), and automatic identification with other chronic respiratory system diseases is carried out; dynamic physiological changes of the airway are captured, continuous monitoring is achieved, and disease comprehensive management for the chronic obstructive pulmonary disease is achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical equipment, and specifically relates to an intelligent dynamic monitoring device and method for chronic obstructive pulmonary disease based on wireless sensing. Background Art

[0002] With the development of the economy and society, the aging of the population, the acceleration of urbanization and industrialization, as well as the change in lifestyle and the prevalence of behavioral risk factors, the prevention and control of chronic non-communicable diseases (referred to as chronic diseases) has become increasingly severe. Deaths caused by chronic diseases have accounted for 73.6% of the total global deaths, becoming the primary problem threatening the health of all mankind in the 21st century. Among the major chronic diseases that threaten the lives and health of residents, chronic obstructive pulmonary disease (COPD) is particularly worthy of attention. According to statistics, the prevalence of COPD in people over 40 years old in my country is as high as 13.7%, with a total number of patients of nearly 100 million, ranking third in the cause of death among Chinese residents. The cost of disease treatment is huge, which brings a heavy burden to patients, their families and society.

[0003] COPD is a heterogeneous lung condition characterized by chronic respiratory symptoms that are persistent, intermittent, atypical, and relatively hidden. Its course is not completely reversible and often develops in a progressive manner. Once the disease occurs, it cannot be cured. Therefore, early screening, early diagnosis, and early intervention are the key to its prevention and control. However, due to the low disease awareness rate, the lack of accurate, efficient, and cost-effective monitoring methods, and the immaturity of standardized follow-up and management technologies, COPD has become the focus and difficulty of chronic disease prevention and control in my country.

[0004] Existing COPD monitoring methods include pulmonary function tests and imaging examinations. However, traditional pulmonary function testers usually need to be operated by professional doctors and performed in hospitals, which is not only time-consuming, but also makes patients feel uncomfortable due to constant blowing. Especially for patients with breathing difficulties, it may be very difficult to perform the blowing test, or even impossible to complete the test, resulting in failure to obtain effective test results. In addition, patients with chest pain or frequent coughing, as well as critically ill patients or unconscious patients may not be able to obtain satisfactory pulmonary function test results. In addition, existing imaging monitoring equipment is often complicated to operate and expensive, requiring patients to follow prescribed steps for testing, and the results rely on the doctor's diagnosis and analysis, with a certain delay, and frequent examinations may be affected by radiation. These devices also require regular inspections and maintenance, which increases costs.

[0005] Currently, there are portable devices developed for primary screening scenarios of primary hospitals and home patients on the market, which can detect the main indicators of lung function. For example, Breathing Home A1 and Orange SP10BT, the main detectable indicators are: FVC, FEV1, FEV1 / FVC, PEF, FEF25%, FEF75%, FEF25-75%. However, the current portable devices still have the following important problems: high requirements for personal operation, the detection method is still mainly based on ventilation function, the detection indicators are limited, the accuracy needs to be verified, the results are relatively broad, and the interpretation needs to be combined with clinical examinations. In response to the above problems, it is urgent to develop a simpler, continuously monitored, and more specialized intelligent dynamic monitor for COPD. Summary of the invention

[0006] Purpose of the invention: In view of the shortcomings of the prior art, the present invention proposes an intelligent dynamic monitoring device and method for COPD based on wireless sensing, which utilizes more effective sensors and wireless sensing and other technical means to make the monitoring method more flexible and the monitoring effect more accurate, and can solve the problems faced by existing COPD monitoring methods.

[0007] Technical solution: The first aspect of the present invention provides a COPD intelligent dynamic monitoring device based on wireless sensing:

[0008] The COPD monitoring device consists of a power module, a data acquisition module, a denoising module, a data processing module, a prediction module, a central processing unit CPU module, a data storage module, a communication module and a user interaction module;

[0009] The power module supplies power to each module to support the operation of the entire COPD monitoring device; the data acquisition module transmits the collected data to the denoising module; the denoising module sends the denoised data to the data processing module for processing, and the data processing module uses an autoencoder to extract data features and adopts a channel attention mechanism for feature fusion; the data processing module is connected to the prediction module to transmit the collected and processed data to the prediction module; there is a unidirectional interface between the CPU module and the prediction module to transmit the prediction results of the COPD by the prediction module to the CPU module; there is a bidirectional interface between the CPU module and the data storage module to transmit all collected data to the data storage module and call data from the data storage module when needed; there is a unidirectional interface between the CPU module and the communication module, and the communication module supports multiple communication protocols so as to be compatible and connected with different types of devices; there is a unidirectional interface between the communication module and the user interaction module to transmit the feedback results and suggestions further generated by the data output by the CPU to the user interaction module.

[0010] Furthermore, the data acquisition module is composed of an electrode sensor unit, a capacitive sensor unit, an acceleration sensor unit, a voice sensor unit, an optical heart rate sensor unit, a breathing sensor unit, a blood oxygen sensor unit, an air quality sensor unit, a signal transmitter unit, and a signal receiver unit, so as to collect data and send the collected multimodal data to the denoising module through the serial port.

[0011] Furthermore, the denoising module denoises the data and constructs a denoising modal autoencoder for each modality, including audio and text, to address the noise interference in the multimodal data of COPD, so as to realize the mining of modal information. Specifically, the denoising modal autoencoder consists of an encoder and a decoder. The encoder encodes the input data into a latent representation, which contains the core information of the input data, thereby projecting the data of each modality into their own lower-dimensional feature spaces. The decoder reconstructs the input of the encoder based on the latent representation.

[0012] Furthermore, the data processing module uses a denoising modal autoencoder to extract features from multi-parameter information for the denoised multimodal data: text data of chest volume changes in images, text data of blood oxygen saturation, and audio data of respiratory sounds. First, the principal component analysis method is used to remove redundant information, reduce data dimensions, and mine the principal components related to audio modality, text modality, and COPD. Then, according to the data type and source, the original data is divided into several independent channels, including a respiratory sound channel, a chest volume change channel, and a heart rate channel. The data of each channel is input into the denoising modal autoencoder, and the feature extraction capability of the autoencoder is used to extract the robust multimodal features of each channel.

[0013] For the multi-channel and multi-modal features generated by the denoising modal autoencoder, a feature-level multimodal fusion model for multi-parameter information is constructed, which comprehensively utilizes information from different modalities to provide a quantitative basis for COPD risk prediction; the multimodal feature outputs extracted by the denoising modal autoencoder are spliced ​​in the channel dimension, and the channel attention mechanism is adopted. Through compression and excitation, the network pays attention to the mutual influence relationship between different channels, learns the importance of each channel feature, extracts the key information in the feature, and performs dimensional transformation of the fully connected layer to achieve feature fusion.

[0014] Furthermore, the prediction module compares and analyzes the feature set formed by feature fusion with the COPD multi-parameter indicator system through the COPD clinical prediction model to obtain the COPD prediction result;

[0015] The COPD clinical prediction model is based on questionnaire survey and physical examination data from a large sample population cohort. It uses LASSO regression for variable screening, and combines kernel function or spline regression system to evaluate the linear and nonlinear effects of variables on outcomes. On this basis, multi-time point evaluation and measurement information are applied, and the exposure trajectory of the research subjects is constructed based on the growth mixture model. The association between changes in the exposure characteristic trajectory and the risk of COPD is evaluated, and finally the model is constructed based on the changes in the trajectory of the screening variables and their key characteristics. The COPD clinical prediction model incorporates multiple indicators of COPD clinical symptoms, clinical signs, and physical function, and has been performance evaluated and model verified to predict the patient's COPD. The prediction of the occurrence of lung disease is a statistical model that can predict the occurrence of outcomes. A case-control study design was adopted to include 500 patients with acute exacerbation of COPD, and stable COPD patients and healthy controls without COPD were matched with propensity scores in a ratio of 1:1 and 1:2, and the key perceptible indicators of COPD of the research subjects were dynamically collected. For multi-time monitoring information of patients with multiple disease courses, the range of changes of indicators and the differences between different groups were analyzed, and the receiver operating characteristic (ROC) curve was used to determine the optimal thresholds of different indicators, and the contribution of different indicators in the identification of COPD was analyzed, so as to construct a multi-parameter indicator system for COPD.

[0016] Furthermore, the user interaction module feeds back the results of the monitoring device to the user in a visual or other form through output devices such as a display and a speaker, and provides lifestyle and medical advice for COPD patients based on relevant data.

[0017] The second aspect of the present invention provides a method for intelligent dynamic monitoring of COPD based on wireless sensing, comprising seven steps:

[0018] S1, the data acquisition module uses electrode sensors, capacitance sensors, acceleration sensors, voice sensors, optical heart rate sensors, breathing sensors, blood oxygen sensors, air quality sensors, signal transmitters, and signal receivers to measure the changes in chest volume and ventilation cycle of COPD patients, collect clinical sign data of COPD patients' respiratory rate, breathing sound, respiratory rhythm, heart rate, heart rhythm, and blood oxygen saturation changes, and obtain gait and somatic function data including stride length and stride frequency, as well as indoor air pollutant data; the collected acceleration data can be used to analyze movement patterns, monitor posture changes, and also for gait recognition;

[0019] S2, the denoising module denoises the data. Aiming at the noise interference in the multimodal data of COPD, a denoising modal autoencoder is constructed for each modality including audio, text, etc. to realize the mining of modal information. The denoising modal autoencoder consists of an encoder and a decoder. The encoder encodes the input data into a potential representation, which contains the core information of the input data, so as to project the data of each modality into their own lower-dimensional feature space. The decoder reconstructs the input of the encoder based on the potential representation. The noisy multimodal data is input into the autoencoder, and the clean input signal is reconstructed to provide a robust potential representation. Data with similar data types are classified into the corresponding channels of the autoencoder, and a specific encoder is used in each channel to process data of similar types.

[0020] S3, the denoising module sends the denoised multimodal data to the data processing module through the serial port. The data processing module uses the denoising modal autoencoder to realize the feature extraction of multi-parameter information;

[0021] S4, the prediction module compares and analyzes the feature set formed by feature fusion with the multi-parameter indicator system of COPD through the COPD clinical prediction model to obtain the COPD prediction result. The COPD clinical prediction model incorporates multiple indicators such as COPD clinical symptoms, clinical signs, and physical function, and has been evaluated and verified by the performance. It can predict the occurrence of COPD in patients and is a statistical model that can predict the occurrence of outcomes;

[0022] S5, the prediction module transmits the prediction result obtained by the COPD analysis to the CPU module, and also transmits to the CPU module the multimodal data collected by the previous data collection module;

[0023] S6, the CPU module sends all data to the data storage module for storage and to the communication module for transmission;

[0024] S7, the communication module establishes a connection with the user interaction module and transmits COPD related data. The user interaction module displays the feedback results on the application interface and gives lifestyle and medical advice for COPD patients based on the relevant data.

[0025] Furthermore, in S1, the capacitive sensor is mainly used to detect touch, distance or the presence of an object. The electrode sensor injects current into the human chest surface and collects voltage signals, and combines the imaging method of parameterized level sets to reconstruct the impedance distribution of the lung organs. The combination of the two can intuitively and continuously display the vibration size of the chest cavity in the form of dynamic images, numerical values ​​and curves to obtain the changes in chest cavity volume and ventilation cycle; set the threshold of the impedance value to distinguish the gas area with low impedance from the non-gas area with high impedance; through threshold segmentation, the boundary of the lung is calibrated to form a three-dimensional impedance image of the lung area.

[0026] Solve the volume v of the three-dimensional image of the lungs, and v = A × h, A is the projection area of ​​the three-dimensional image of the lungs, first solve the average height h of the image, the average height is calculated by the flattening rate, the formula is h = λ × D min , where λ refers to the parameter related to the flattening rate; the volume of the three-dimensional image of the lung is determined as

[0027] v=A×λ×D min (1)

[0028] A multivariate linear regression model with λ as a function is established. The model takes the characteristic parameters in the three-dimensional image as variables. The model is as follows

[0029] λ=α0+α1ρ+α2Φ+α3θ (2)

[0030] Among them, ρ, Φ, and θ represent the elongation, roundness, and edge smoothness of the three-dimensional lung image, respectively, and α0, α1, α2, and α3 are the coefficients to be determined;

[0031]

[0032] The elongation ρ is the major diameter D max With short diameter D min A and P are the projection area and projection perimeter of the lung 3D image, respectively; edge smoothness θ is the square of the variance s of the marker line and the mean R of the marker line arg 2 The ratio of the identification line average R arg is the average length from the centroid of the lung image to its edge;

[0033] For the existing multiple three-dimensional images with known volumes, the flattening rate is inverted using formula (1), and the flattening rate and corresponding characteristic parameters of each of the four three-dimensional images are substituted into formula (2) to obtain a set of model coefficients. In this way, multiple sets of coefficients are obtained. For different sets of coefficients, the least square error method in mathematical statistics is used to determine the regression coefficient. Once formula (2) is determined, the flattening rate of the three-dimensional image of the same lung to be tested is obtained using formula (2), and then the volume is estimated using formula (1).

[0034] The voice sensor is used to capture sound signals for chest breathing sound recognition, environmental sound monitoring, etc. The voice sensor can obtain the chest sound of the patient when breathing, and then judge the condition of COPD; the data collected by the optical heart rate sensor is used to analyze the heart rate; the breathing sensor is used to obtain the breathing rate and breathing rhythm of COPD patients; the blood oxygen sensor and air quality sensor can obtain the patient's blood oxygen saturation changes and indoor air pollutant data.

[0035] Furthermore, in S2, for the i-th channel, the input noisy data X i ={Xi,1 ,X i,2 ,…,X i,N}, the process through the encoder is:

[0036]

[0037] Among them, θ i is the learning parameter of the encoder, N is the total number of samples in the i-th channel, h i is the potential representation of the sample;

[0038] The potential representation is fed into the decoder, and the process through the decoder is

[0039]

[0040] in, is the learning parameter of the decoder, To reconstruct samples; the mean square error function is used during training

[0041]

[0042] in, is the regularization factor; the reconstructed sample obtained This is the denoised data.

[0043] Furthermore, in S3, first, through principal component analysis, the multimodal data of m types and n samples are expressed by matrix X as follows:

[0044]

[0045] Next, data standardization

[0046]

[0047] Among them, u is the average value, S j is the standard deviation;

[0048] Calculate the standardized data correlation coefficient matrix R:

[0049]

[0050] where r ij is the correlation coefficient between the i-th row and j-th column of the standardized matrix Z;

[0051] Solve for the eigenvalues ​​of the matrix R; use the Jacobian method | R-βI p |=0 to find the eigenvalue β of the correlation coefficient matrix R i , sort the eigenvalues ​​from large to small; I p is the identity matrix;

[0052] Calculate the proportion of each eigenvalue to the total number of eigenvalues, that is, the variance contribution rate:

[0053]

[0054] in, For the cumulative variance contribution rate, select the component F corresponding to the eigenvalue with the larger variance contribution rate C As the cth principal component; reconstruct the multimodal data after removing redundant information and reducing the dimension by c principal components, and mine the principal components related to audio modality, text modality and COPD; then, according to the data type and source, divide the original data into several independent channels, including breathing sound channel, chest volume change channel, and heart rate channel; the data of each channel is input into the denoising modality autoencoder respectively, and the feature extraction ability of the autoencoder is used to extract the robust multimodal features of each channel;

[0055] The number of unlabeled samples input to the autoencoder is M, the feature dimension is D, the number of hidden layers of the autoencoder is L, and the number of neurons in the top layer of the hidden layer is S L , the weight of the autoencoder unit’s contribution to the final output features:

[0056]

[0057] The bias unit vector of each hidden layer constitutes the bias unit matrix of the model:

[0058]

[0059] Input sample vector The nonlinear representation in the lth hidden layer of the autoencoder is And the input layer l = 0, then

[0060]

[0061] The decoding process is actually the inverse process of encoding. The weight coefficient of the neuron propagation path is symmetrical with the encoding process. The decoding reconstruction of the output layer l = 2L is calculated, where L is the number of top layers of the hidden layer:

[0062]

[0063] The two important encoding parameter vectors of the communication radiation source signal sample in the autoencoder are recorded as a training parameter matrix:

[0064] W=[(ω (1) ,b (1) ),(ω (2) ,b (2) )…,(ω (l) ,b (l) )]T (17)

[0065] First, construct the objective function as:

[0066]

[0067] Among them, the reconstruction matrix Y is the self-encoder according to the encoding coefficient ω (1) and bias unit b (1) For the input sample matrix X U Reconstruction; γ is the weight attenuation coefficient to avoid overfitting of the objective function to the sample; s l Indicates that the current autoencoder is the sth l Layer; weight coefficient vector is the weight coefficient of the propagation path between the jth neuron in the lth layer and the ith neuron in the l+1th layer;

[0068] Use the iterative optimal value method to solve the training parameter W that minimizes the objective function;

[0069] The gradient of J(W) on the jth neuron in the lth layer of the network hidden layer is It constitutes the gradient matrix Perform Taylor second-order expansion of J(W) (the subscript k+1 indicates the number of iterations):

[0070]

[0071] Derivation

[0072] g(W)=g(W k+1 )+H k+1 (WW k+1 ) (20)

[0073] Among them, H k+1 is the k+1th calculation result of the Hessian matrix of the objective function; generally, H0 required for the first iteration is set to be the unit diagonal matrix, and the kth iteration value of W is recorded as W k , and continue to transform into

[0074]

[0075] In the above formula, let the training parameter change matrix

[0076] s k =W k+1 -W k (twenty two)

[0077] Gradient change matrix

[0078] y k =g k+1 -g k(twenty three)

[0079] B k+1 Represents the inverse of the Hessian matrix of the objective function to be estimated but

[0080] B k+1 ·y k =s k (twenty four)

[0081] Let the auxiliary operator

[0082]

[0083] In the above process, s k With y k All depend on the training parameters required

[0084] W pre =[ω pre ,b pre ] (26)

[0085] ω pre is the weight of each hidden layer of the autoencoder contributing to the final output feature, b pre is the bias unit vector of each hidden layer.

[0086] Each B in the k+1 iteration formula k+1 , solve the sample feature W pre ∈R D×N , the training parameters extracted at this time are the desired features;

[0087] The multimodal feature outputs extracted by the denoising modal autoencoder are spliced ​​in the channel dimension, and the channel attention mechanism is adopted. Through compression and excitation, the network pays attention to the mutual influence relationship between different channels, learns the importance of each channel feature, extracts the key information in the feature, and performs dimensional transformation of the fully connected layer to achieve feature fusion; the compression and excitation network of the channel attention mechanism is mainly composed of compression units and excitation units; the compression unit uses the pooling layer to sum and compress the global spatial features in each channel to form their own channel features; pooling divides the feature space into multiple subspaces, selects a value in each subspace, and forms a new feature space with the selected values ​​to reduce the dimension of the feature space; the excitation unit uses two fully connected layers to obtain the weight parameters w1 and w2 of each layer to reduce the number of parameters of the unit and enhance the migration ability of the unit; during the training process of the model, the feature weights of each channel domain and the correlation between channels are obtained; the calculation formula of the channel attention mechanism is as follows:

[0088]

[0089] u c =σ(w2δ(w1u s )) (28)

[0090]

[0091] Where σ is the summation function. δ represents the delta function, also known as the impulse function. Scale is the dimension transformation. u s is the normalized global information of the concatenated output, T is the time step, u c is the weight corresponding to each channel, z(i) represents the output result after splicing, It is the weighted output; finally, a dimensional transformation of the fully connected layer is performed to achieve feature fusion.

[0092] Compared with the traditional COPD monitoring device and method, the beneficial effects of the present invention are:

[0093] (1) Simplification: Get rid of the reliance on breath testing, reduce the difficulty of testing, and use other signal monitoring methods to reflect the severity and variability of airway obstruction.

[0094] (2) Multiple indicators: Comprehensive assessment based on extrapulmonary indicators (such as lung volume changes, heart rate, blood oxygen, breath sounds, gait, etc.), and automatic differentiation from other chronic respiratory diseases.

[0095] (3) Continuity: Capture dynamic physiological changes in the airway, achieve continuous monitoring, and provide comprehensive disease management for COPD. BRIEF DESCRIPTION OF THE DRAWINGS

[0096] Figure 1 This is a module composition diagram of the COPD intelligent dynamic monitoring device based on wireless sensing in an embodiment of the present invention.

[0097] Figure 2 It is a data acquisition module diagram in an embodiment of the present invention.

[0098] Figure 3 It is a diagram of a data processing module and a prediction module in an embodiment of the present invention.

[0099] Figure 4 This is a flow chart of the intelligent dynamic monitoring method for chronic obstructive pulmonary disease based on wireless sensing in an embodiment of the present invention.

[0100] Figure 5 This is a diagram of an intelligent dynamic monitoring device for chronic obstructive pulmonary disease based on wireless sensing in an embodiment of the present invention.

[0101] Numbers in the accompanying drawings: 1-user interaction device, 2-built-in battery, 3-PCB integrated circuit board, 4-communication interface, 5-memory, 6-microprocessor chip, 7-charging interface, 8-user interaction device communication interface, 9-user interaction device charging interface, 10-charging indicator light, 11-signal transmitter, 12-signal receiver. DETAILED DESCRIPTION

[0102] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings.

[0103] like Figure 1 As shown, the COPD monitoring device is composed of a power module, a data acquisition module, a denoising module, a data processing module, a prediction module, a central processing unit (CPU) module, a data storage module, a communication module and a user interaction module. The data acquisition module transmits the collected data to the denoising module; the denoising module sends the denoised data to the data processing module for processing; there is a one-way interface between the CPU module and the data processing module to realize the transmission of various collected and processed data to the CPU module; there is a two-way interface between the CPU module and the data storage module to realize the transmission of all collected data to the data storage module and the call of data from the data storage module when needed; there is a one-way interface between the CPU module and the communication module, and the communication module supports multiple communication protocols so as to be compatible and connected with different types of devices; there is a one-way interface between the communication module and the user interaction module to realize the transmission of the data output by the CPU and the feedback results and suggestions generated to the user interaction module.

[0104] Among them, the power supply module supplies power to each module of the entire COPD monitoring device to support the operation of the device.

[0105] The data acquisition module consists of an electrode sensor unit, a capacitive sensor unit, an acceleration sensor unit, a voice sensor unit, an optical heart rate sensor unit, a respiratory sensor unit, a blood oxygen sensor unit, an air quality sensor unit, a signal transmitter unit, and a signal receiver unit. It is used to measure the changes in chest volume and ventilation cycle of patients with COPD, collect clinical signs data such as respiratory rate, respiratory sound, respiratory rhythm, heart rate, heart rhythm, and blood oxygen saturation changes of patients with COPD, and obtain gait: stride, step frequency and other physical function data, as well as indoor air pollutant (PM2.5, inhalable particulate matter, etc.) data.

[0106] Exemplarily, the data acquisition module uses electrode sensors, acceleration sensors, capacitance sensors, voice sensors, optical heart rate sensors, signal transmitters, signal receivers, etc. to collect data. The collected acceleration data can be used to analyze movement patterns, monitor posture changes, and can also be used for gait recognition. Capacitive sensors are mainly used to detect touch, distance or the presence of objects. The electrode sensor injects current into the surface of the human chest and collects voltage signals, combined with the imaging method of parameterized level sets, to reconstruct the impedance distribution of the lung organs. The combination of the two can intuitively and continuously display the vibration size of the chest cavity in the form of dynamic images, numerical values ​​and curves to obtain changes in chest cavity volume and ventilation cycles. The voice sensor is used to capture sound signals and can be used for voice recognition, environmental sound monitoring, etc. The sound of the patient's breathing can be obtained through the voice sensor. In addition, the data collected by the optical heart rate sensor can be used to analyze the heart rate. The respiratory sensor can be used to obtain the respiratory rate and respiratory rhythm of patients with chronic obstructive pulmonary disease. Among them, the heart rate and respiratory rate can be measured by optical heart rate sensors and respiratory sensors, or by transmitting wifi signals through a signal transmitter, receiving wifi signals through a signal receiver, and then analyzing the wifi signals. The blood oxygen sensor and air quality sensor can obtain the patient's blood oxygen saturation changes and indoor air pollutant data. Finally, the collected multimodal data is sent to the denoising module through the serial port.

[0107] The denoising module denoises the data. Aiming at the noise interference in the multimodal data of COPD, a denoising modal autoencoder is constructed for each modality including audio, text, etc. to realize the mining of modal information. Specifically, the denoising modal autoencoder consists of an encoder and a decoder. The encoder encodes the input data into a potential representation, which contains the core information of the input data, thereby projecting the data of each modality into their own lower-dimensional feature space. The decoder reconstructs the input of the encoder based on the potential representation. The noisy multimodal data is input into the autoencoder, and the clean input signal is reconstructed to provide a robust potential representation. Data with similar data types, such as audio data such as breathing sounds and text data such as heart rate and respiratory rate, are classified into multiple channels of the autoencoder. Therefore, a specific encoder is used in each channel to process similar types of data.

[0108] For the i-th channel, the input noisy data X i ={X i,1 ,X i,2 ,…,X i,N}, the process through the encoder is

[0109]

[0110] Among them, θ i is the learnable parameter of the encoder, N is the total number of samples in the i-th channel, and hi is the potential representation of the sample.

[0111] The potential representation is fed into the decoder, and the process through the decoder is

[0112]

[0113] in, are the learnable parameters of the decoder, is the reconstructed sample. The mean square error function is used during training

[0114]

[0115] in, is the regularization factor. The reconstructed sample obtained This is the denoised data.

[0116] The data processing module uses the denoising modal autoencoder to extract features from multi-parameter information for denoised multimodal data: chest volume changes, text data such as blood oxygen saturation, and audio data such as breathing sounds. First, the principal component analysis method is used to remove redundant information, reduce data dimensions, and mine the principal components related to audio modality, text modality, and COPD. Then, according to the data type and source, the original data is divided into several independent channels, such as breathing sound channel, chest volume change channel, heart rate channel, etc. The data of each channel is input into the denoising modal autoencoder separately, and the feature extraction capability of the autoencoder is used to extract the robust multimodal features of each channel.

[0117] For the multi-channel multimodal features generated by the denoising modal autoencoder, a feature-level multimodal fusion model for multi-parameter information is proposed, which comprehensively utilizes the information of different modalities to provide a quantitative basis for COPD risk prediction. The multimodal feature output extracted by the denoising modal autoencoder is spliced ​​in the channel dimension, and the channel attention mechanism is adopted. Through compression and excitation, the network pays attention to the mutual influence relationship between different channels, and learns the importance of each channel feature at the same time, effectively extracting the key information in the feature. The channel attention mechanism-compression and excitation network is mainly composed of compression units and excitation units. The compression unit uses the pooling layer to sum and compress the global spatial features in each channel to form their own channel features. Pooling divides the feature space into multiple subspaces, selects a value in each subspace, and forms a new feature space with the selected values, which can reduce the dimension of the feature space. The excitation unit uses two fully connected layers to obtain the weight parameters w1 and w2 of each layer to reduce the number of parameters of the unit and enhance the migration ability of the unit. During the training process of the model, the feature weights of each channel domain and the correlation between channels can be obtained. The calculation formula of the channel attention mechanism is as follows:

[0118]

[0119] u c =σ(w2δ(w1u s )) (28)

[0120]

[0121] Where σ is the summation function. δ represents the delta function, also known as the impulse function. Scale is the dimension transformation. u s is the normalized global information of the concatenated output, T is the time step, u c is the weight corresponding to each channel, z represents the output result after splicing, is the weighted output. Finally, a dimension transformation of the fully connected layer is performed to achieve feature fusion.

[0122] The prediction module compares and analyzes the feature set formed by feature fusion with the expert knowledge base through the COPD clinical prediction model to obtain the COPD prediction results. The COPD clinical prediction model incorporates multiple indicators such as COPD clinical symptoms, clinical signs, and physical function, and after performance evaluation and model verification, it can predict the occurrence of COPD in patients. It is a statistical model that can predict the occurrence of outcomes.

[0123] The CPU module is a main control module based on a microprocessor chip, which realizes the coordinated operation between the various modules of the entire device. Specifically, the CPU module receives data analyzed by the data acquisition module and the prediction module, including chest volume change data, respiratory rate, heart rate, and the user's disease prediction result data. After receiving all the data, the CPU module sends all the data to the data storage module for storage and to the communication module for transmission.

[0124] The data storage module is used to store all the data acquired by the CPU module, and use the CPU module to transfer all the collected data to the data storage module and call data from the data storage module when needed.

[0125] The communication module supports multiple communication protocols so that different types of user interaction devices can be compatible and connected with the monitoring device;

[0126] Exemplarily, the communication module uses existing communication technology to transmit data, uses the communication protocol to establish a connection with the user interaction device and transmit COPD related data.

[0127] The user interaction module can provide the user with visual or other feedback from the monitoring device, such as through output devices such as displays and speakers, and provide lifestyle and medical advice for COPD patients based on relevant data.

[0128] like Figure 2 As shown, the data acquisition module uses electrode sensors, acceleration sensors, capacitance sensors, voice sensors, optical heart rate sensors, signal transmitters, signal receivers, etc. to collect data. The collected acceleration data can be used to analyze movement patterns, monitor posture changes, and can also be used for gait recognition. Capacitive sensors are mainly used to detect touch, distance or the presence of objects. The electrode sensor injects current into the human chest surface and collects voltage signals, combined with the imaging method of parameterized level sets, to reconstruct the impedance distribution of the lung organs. The combination of the two can intuitively and continuously display the vibration size of the chest cavity in the form of dynamic images, numerical values ​​and curves to obtain changes in chest cavity volume and ventilation cycles. The voice sensor is used to capture sound signals and can be used for voice recognition, environmental sound monitoring, etc. The voice sensor can obtain the sound of the patient's breathing. In addition, the data collected by the optical heart rate sensor can be used to analyze the heart rate. The respiratory sensor can be used to obtain the respiratory rate and respiratory rhythm of patients with chronic obstructive pulmonary disease. Among them, the heart rate and respiratory rate can be measured by optical heart rate sensors and respiratory sensors, or by transmitting wifi signals by the signal transmitter, receiving wifi signals by the signal receiver, and then analyzing the wifi signals. Blood oxygen sensors and air quality sensors can obtain data on changes in patient blood oxygen saturation and indoor air pollutants.

[0129] like Figure 3 As shown in the figure, the data processing module uses the denoising modal autoencoder to extract features of multi-parameter information for the denoised multimodal data: chest volume changes, blood oxygen saturation and other text data, and audio data such as breathing sounds. First, the principal component analysis method is used to remove redundant information, reduce data dimensions, and mine the principal components related to audio modality, text modality and COPD. Then, according to the data type and source, the original data is divided into several independent channels, such as breathing sound channel, chest volume change channel, heart rate channel, etc. The data of each channel is input into the denoising modal autoencoder respectively, and the robust multimodal features of each channel are extracted by using the feature extraction capability of the autoencoder.

[0130] For the multi-channel multi-modal features generated by the denoising modal autoencoder, a feature-level multi-modal fusion model for multi-parameter information is proposed, which comprehensively utilizes information from different modalities to provide a quantitative basis for COPD risk prediction. The multi-modal feature output extracted by the denoising modal autoencoder is spliced ​​in the channel dimension, and the channel attention mechanism is adopted. Through compression and excitation, the network pays attention to the mutual influence relationship between different channels, and at the same time learns the importance of each channel feature, effectively extracts the key information in the feature, and performs dimensional transformation of the fully connected layer to achieve feature fusion.

[0131] The prediction module uses the COPD clinical prediction model to compare and analyze the feature set formed by feature fusion with the COPD multi-parameter indicator system to obtain the COPD prediction results. The COPD clinical prediction model incorporates multiple indicators such as COPD clinical symptoms, clinical signs, and physical function, and after performance evaluation and model verification, it can predict the occurrence of COPD in patients. It is a statistical model that can predict the occurrence of outcomes.

[0132] like Figure 4 As shown, the second aspect of the present invention provides a method for intelligent dynamic monitoring of chronic obstructive pulmonary disease, comprising the following steps:

[0133] S1-1, the data acquisition module uses electrode sensors, capacitance sensors, acceleration sensors, voice sensors, optical heart rate sensors, breathing sensors, blood oxygen sensors, air quality sensors, signal transmitters, and signal receivers to measure the changes in chest volume and ventilation cycle of COPD patients, collect clinical signs data such as respiratory rate, breathing sound, respiratory rhythm, heart rate, heart rhythm, and blood oxygen saturation changes of COPD patients, and obtain gait: stride, step frequency and other physical function data, as well as indoor air pollutants (PM2.5, inhalable particulate matter, etc.) data. The collected multimodal data is sent to the denoising module through the serial port. Among them, the heart rate and respiratory rate can be measured by optical heart rate sensors and breathing sensors, or by transmitting wifi signals through signal transmitters, receiving wifi signals through signal receivers, and then analyzing wifi signals.

[0134] S1-2, the denoising module denoises the data. Aiming at the noise interference in the multimodal data of COPD, a denoising modal autoencoder is constructed for each modality including audio, text, etc. to realize the mining of modal information. Specifically, the denoising modal autoencoder consists of an encoder and a decoder. The encoder encodes the input data into a potential representation, which contains the core information of the input data, thereby projecting the data of each modality into their own lower-dimensional feature space. The decoder reconstructs the input of the encoder based on the potential representation. The noisy multimodal data is input into the autoencoder, and the clean input signal is reconstructed to provide a robust potential representation.

[0135] S1-3, the denoising module sends the denoised multimodal data to the data processing module through the serial port. The data processing module uses the denoising modal autoencoder to realize the feature extraction of multi-parameter information. First, the principal component analysis method is used to remove redundant information, reduce the data dimension, and mine the principal components related to audio modality, text modality and COPD. Then, according to the data type and source, the original data is divided into several independent channels, such as respiratory sound channel, chest volume change channel, heart rate channel, etc. The data of each channel is input into the denoising modal autoencoder respectively, and the robust multimodal features of each channel are extracted by using the feature extraction ability of the autoencoder. The multimodal feature output extracted by the denoising modal autoencoder is spliced ​​in the channel dimension, and the channel attention mechanism is adopted. Through compression and excitation, the network pays attention to the mutual influence relationship between different channels, and at the same time learns the importance of each channel feature, effectively extracts the key information in the feature, and performs dimensional transformation of the fully connected layer to achieve feature fusion.

[0136] S1-4, the prediction module compares and analyzes the feature set formed by feature fusion with the multi-parameter indicator system of COPD through the COPD clinical prediction model to obtain the COPD prediction results. The COPD clinical prediction model incorporates multiple indicators such as COPD clinical symptoms, clinical signs, and physical function, and after performance evaluation and model verification, it can predict the occurrence of COPD in patients. It is a statistical model that can predict the occurrence of outcomes.

[0137] S1-5, the prediction module transmits the prediction results obtained from the COPD analysis to the CPU module, and also transmits to the CPU module the multimodal data collected by the previous data acquisition module.

[0138] S1-6, the CPU module sends all data to the data storage module for storage and to the communication module for transmission.

[0139] S1-7, the communication module establishes a connection with the user interaction module and transmits COPD-related data. The user interaction module displays the feedback results on the application interface and gives lifestyle and medical advice for COPD patients based on the relevant data.

[0140] like Figure 5As shown, a diagram of an intelligent dynamic monitoring device for chronic obstructive pulmonary disease based on wireless sensing of the present invention is shown. The main body of the device is in the shape of a long strip, and three chambers are formed inside the main body, which are respectively used to place a built-in battery 2, a memory 5 and a microprocessor chip 6. The built-in battery 2 can be charged and discharged, and the microprocessor chip 6 can perform data processing and prediction. The signal transmitter 11 and the signal receiver 12 are located on the left and right sides of the device, and are used to transmit and receive wifi signals. Two PCB integrated circuit boards 3 are respectively located in the chambers at both ends. The low-power PCB integrated circuit board 3 integrates multiple sensors and control circuits into a miniaturized circuit board, which reduces the overall volume and weight of the device, and reasonably packages the sensor chips carried to ensure the durability and stability of the device. The chamber is closed by a cover. There is a communication interface 4 on the side of the device, which can be connected to the user interaction device 1. In addition, the PCB integrated circuit board 3, the memory 5 and the microprocessor chip 6 are all connected to the built-in battery 2 through a connecting line, and the built-in battery 2 supplies power to each device of the entire chronic obstructive pulmonary disease monitoring device to support the operation of the device. The connecting line is embedded between the skeleton and the body.

[0141] The PCB integrated circuit board 3, signal transmitter 11 and signal receiver 12 of the device are connected to the built-in battery 2 and microprocessor chip 6 through connecting wires. The PCB integrated circuit board 3 includes an electrode sensor chip, a capacitance sensor chip, an acceleration sensor chip, a voice sensor chip, an optical heart rate sensor chip, a breathing sensor chip, a blood oxygen sensor chip, and an air quality sensor chip, which are used to measure the changes in chest volume and ventilation cycle of COPD patients, collect clinical physical sign data such as respiratory rate, breathing sound, respiratory rhythm, heart rate, heart rhythm, and blood oxygen saturation changes of COPD patients, and obtain gait: stride, step frequency and other body function data, as well as indoor air pollutant (PM2.5, inhalable particulate matter, etc.) data. At the same time, the collected data is transmitted to the microprocessor chip 6. Among them, the heart rate and respiratory rate can be measured by the optical heart rate sensor chip and the breathing sensor chip, or by the signal transmitter 11 transmitting a wifi signal, the signal receiver 12 receiving the wifi signal, and then analyzing the wifi signal to measure, and the collected data is transmitted to the microprocessor chip 6.

[0142] Microprocessor chip 6 is implanted with an artificial intelligence algorithm library to perform data denoising on the collected data, extract features such as threshold, fluctuation, amplitude, frequency, etc. from the data collected by the sensor, and use the channel attention mechanism to fuse the features of text and voice data. In addition, by deploying a lightweight COPD clinical prediction model in microprocessor chip 6, a joint analysis of a multi-parameter indicator system is performed to achieve a millisecond-level ultra-low latency response for the prediction task and obtain COPD prediction results.

[0143] The microprocessor chip 6 sends all data to the memory 5 for storage and transmits it through the communication interface 4. The microprocessor chip 6 is used to transmit all collected data to the memory 5 for storage and to call data from the memory 5 when needed. The user interaction device 1 can be connected to the user interaction device communication interface 8 through the communication interface 4 using a data line, and the data of the monitoring device can be fed back to the user in a visual or other form, such as through output devices such as a display and a speaker. And lifestyle and medical advice for COPD patients are given based on relevant data.

[0144] Plug in the charging cable, connect it to the charging interface 7 and the user interaction device charging interface 9, start charging the built-in battery 2 of the device and the user interaction device 1, observe the charging indicator light 10 and the power displayed on the user interaction device screen, and you can know the charging status of the built-in battery 2 and the user interaction device 1. When using, place the device on the chest and connect it to the user interaction device 1 with a data cable. Press the menu key on the user interaction device 1, select the sensor mode or the wifi mode to measure various data of COPD and obtain the prediction results of COPD. Among them, the heart rate and respiratory rate in the sensor mode can be measured by the optical heart rate sensor chip and the respiratory sensor chip, and the heart rate and respiratory rate in the wifi mode can be measured by the signal transmitter 11 transmitting the wifi signal, the signal receiver 12 receiving the wifi signal, and then analyzing the wifi signal.

[0145] The above description is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiment. Any equivalent modifications or changes made by ordinary technicians in this field based on the contents disclosed by the present invention should be included in the protection scope recorded in the claims.

Claims

1. An intelligent dynamic monitoring device for chronic obstructive pulmonary disease based on wireless sensing, characterized in that: The COPD monitoring device is composed of a power module, a data acquisition module, a denoising module, a data processing module, a prediction module, a central processing unit CPU module, a data storage module, a communication module and a user interaction module; The power module supplies power to each module to support the operation of the entire COPD monitoring device; the data acquisition module transmits the collected data to the denoising module; the denoising module sends the denoised data to the data processing module for processing, and the data processing module uses the autoencoder to extract data features and adopts the channel attention mechanism for feature fusion; the data processing module is connected to the prediction module to realize the transmission of various collected and processed data to the prediction module; there is a unidirectional interface between the CPU module and the prediction module to realize the transmission of the prediction result of the COPD by the prediction module to the CPU module; there is a bidirectional interface between the CPU module and the data storage module to realize the transmission of all collected data to the data storage module and the call of data from the data storage module when needed; There is a one-way interface between the CPU module and the communication module. The communication module supports multiple communication protocols so as to be compatible and connected with different types of devices. There is a one-way interface between the communication module and the user interaction module to realize the transmission of feedback results and suggestions generated by the data output by the CPU to the user interaction module.

2. The COPD intelligent dynamic monitoring device based on wireless sensing according to claim 1 is characterized by: The data acquisition module is composed of an electrode sensor unit, a capacitive sensor unit, an acceleration sensor unit, a voice sensor unit, an optical heart rate sensor unit, a breathing sensor unit, a blood oxygen sensor unit, an air quality sensor unit, a signal transmitter unit, and a signal receiver unit, so as to collect data and send the collected multimodal data to the denoising module through the serial port.

3. The COPD intelligent dynamic monitoring device based on wireless sensing according to claim 1 is characterized by: The denoising module denoises the data, and constructs a denoising modal autoencoder for each modality including audio and text to address the noise interference in the multimodal data of COPD, thereby realizing the mining of modal information; specifically, the denoising modal autoencoder consists of an encoder and a decoder. The encoder encodes the input data into a potential representation, which contains the core information of the input data, thereby projecting the data of each modality into their own lower-dimensional feature spaces. The decoder reconstructs the input of the encoder based on the potential representation.

4. The COPD intelligent dynamic monitoring device based on wireless sensing according to claim 1 is characterized by: The data processing module uses a denoising modal autoencoder to extract features of multi-parameter information from the denoised multimodal data: text data of chest volume changes in images, text data of blood oxygen saturation, and audio data of breathing sounds. First, the principal component analysis method is used to remove redundant information, reduce data dimensions, and mine the principal components related to audio modality, text modality, and COPD. Then, according to the data type and source, the original data is divided into several independent channels, including the breathing sound channel, the chest volume change channel, and the heart rate channel. The data of each channel is input into the denoising modal autoencoder respectively, and the feature extraction capability of the autoencoder is used to extract the robust multimodal features of each channel. For the multi-channel and multi-modal features generated by the denoising modal autoencoder, a feature-level multimodal fusion model for multi-parameter information is constructed, which comprehensively utilizes information from different modalities to provide a quantitative basis for COPD risk prediction; the multimodal feature outputs extracted by the denoising modal autoencoder are spliced ​​in the channel dimension, and the channel attention mechanism is adopted. Through compression and excitation, the network pays attention to the mutual influence relationship between different channels, learns the importance of each channel feature, extracts the key information in the feature, and performs dimensional transformation of the fully connected layer to achieve feature fusion.

5. The COPD intelligent dynamic monitoring device based on wireless sensing according to claim 1 is characterized by: The prediction module compares and analyzes the feature set formed by feature fusion with the COPD multi-parameter indicator system through the COPD clinical prediction model to obtain the COPD prediction result; The COPD clinical prediction model is based on questionnaire survey and physical examination data of a large sample population cohort, uses LASSO regression for variable screening, and combines kernel function or spline regression system to evaluate the linear and nonlinear effects of variables on outcomes. On this basis, multi-time point evaluation and measurement information are applied, and the exposure trajectory of the research subjects is constructed based on the growth mixture model, and the association between changes in exposure characteristic trajectories and the risk of COPD is evaluated. The model is finally constructed based on the changes in the trajectories of the screening variables and their key characteristics. The COPD clinical prediction model incorporates multiple indicators of COPD clinical symptoms, clinical signs, and physical function, and has been evaluated for performance and model validation. The prediction of the occurrence of COPD is a statistical model that can predict the occurrence of outcomes. A case-control study design was adopted to include 500 patients with acute exacerbation of COPD, and stable COPD patients and healthy controls without COPD were matched with propensity scores in a ratio of 1:1 and 1:2, and the key perceptible indicators of COPD of the research subjects were dynamically collected. The multi-time monitoring information of patients with multiple disease courses was used to analyze the range of changes in indicators and the differences between different groups. The receiver operating characteristic (ROC) curve was used to determine the optimal thresholds of different indicators, and the contribution of different indicators in the identification of COPD was analyzed to construct a multi-parameter indicator system for COPD.

6. The COPD intelligent dynamic monitoring device based on wireless sensing according to claim 1 is characterized by: The user interaction module feeds back the results of the monitoring device to the user in a visual or other form through the output devices of a display and a speaker, and provides lifestyle and medical advice for COPD patients based on the relevant data.

7. A monitoring method using the COPD intelligent dynamic monitoring device based on wireless sensing as described in any one of claims 1 to 6, characterized in that: The method comprises the following steps: S1, the data acquisition module uses electrode sensors, capacitance sensors, acceleration sensors, voice sensors, optical heart rate sensors, breathing sensors, blood oxygen sensors, air quality sensors, signal transmitters, and signal receivers to measure the changes in chest volume and ventilation cycle of COPD patients, collect clinical sign data of COPD patients' respiratory rate, breathing sound, respiratory rhythm, heart rate, heart rhythm, and blood oxygen saturation changes, and obtain gait and somatic function data including stride length and stride frequency, as well as indoor air pollutant data; the collected acceleration data can be used to analyze movement patterns, monitor posture changes, and also for gait recognition; S2, the denoising module denoises the data. Aiming at the noise interference in the multimodal data of COPD, a denoising modal autoencoder is constructed for each modality including audio, text, etc. to realize the mining of modal information. The denoising modal autoencoder consists of an encoder and a decoder. The encoder encodes the input data into a potential representation, which contains the core information of the input data, so as to project the data of each modality into their own lower-dimensional feature space. The decoder reconstructs the input of the encoder based on the potential representation. The noisy multimodal data is input into the autoencoder, and the clean input signal is reconstructed to provide a robust potential representation. Data with similar data types are classified into the corresponding channels of the autoencoder, and a specific encoder is used in each channel to process data of similar types. S3, the denoising module sends the denoised multimodal data to the data processing module through the serial port. The data processing module uses the denoising modal autoencoder to realize the feature extraction of multi-parameter information; S4, the prediction module compares and analyzes the feature set formed by feature fusion with the multi-parameter indicator system of COPD through the COPD clinical prediction model to obtain the COPD prediction result. The COPD clinical prediction model incorporates multiple indicators such as COPD clinical symptoms, clinical signs, and physical function, and has been evaluated and verified by the performance. It can predict the occurrence of COPD in patients and is a statistical model that can predict the occurrence of outcomes; S5, the prediction module transmits the prediction result obtained by the COPD analysis to the CPU module, and also transmits to the CPU module the multimodal data collected by the previous data collection module; S6, the CPU module sends all data to the data storage module for storage and to the communication module for transmission; S7, the communication module establishes a connection with the user interaction module and transmits COPD related data. The user interaction module displays the feedback results on the application interface and gives lifestyle and medical advice for COPD patients based on the relevant data.

8. The method for intelligent dynamic monitoring of COPD based on wireless sensing according to claim 7 is characterized by: In S1, capacitive sensors are mainly used to detect touch, distance or the presence of objects. The electrode sensor injects current into the human chest surface and collects voltage signals, combined with the imaging method of parameterized level sets to reconstruct the impedance distribution of the lung organs. The combination of the two can intuitively and continuously display the vibration size of the chest cavity in the form of dynamic images, numerical values ​​and curves to obtain the changes in chest cavity volume and ventilation cycle; set the threshold of the impedance value to distinguish the gas area with low impedance from the non-gas area with high impedance; through threshold segmentation, the boundary of the lung is calibrated to form a three-dimensional impedance image of the lung area. Solve the volume v of the three-dimensional image of the lungs, and v = A × h, A is the projection area of ​​the three-dimensional image of the lungs, first solve the average height h of the image, the average height is calculated by the flattening rate, the formula is h = λ × D min , where λ refers to the parameter related to the flattening rate; the volume of the three-dimensional image of the lung is determined as v=A×λ×D min (1) A multivariate linear regression model with λ as a function is established. The model takes the characteristic parameters in the three-dimensional image as variables. The model is as follows λ=ɑ0+ɑ1ρ+ɑ2Φ+ɑ3θ(2) Among them, ρ, Φ, and θ represent the elongation, roundness, and edge smoothness of the three-dimensional lung image, respectively, and ɑ0, ɑ1, ɑ2, and ɑ3 are the coefficients to be determined; The elongation ρ is the major diameter D max With short diameter D min A and P are the projection area and projection perimeter of the lung 3D image, respectively; edge smoothness θ is the square of the variance s of the marker line and the mean R of the marker line arg 2 The ratio of the identification line average R arg is the average length from the centroid of the lung image to its edge; For the existing multiple three-dimensional images with known volumes, the flattening rate is inverted using formula (1), and the flattening rate and corresponding characteristic parameters of each of the four three-dimensional images are substituted into formula (2) to obtain a set of model coefficients. In this way, multiple sets of coefficients are obtained. For different sets of coefficients, the least square error method in mathematical statistics is used to determine the regression coefficient. Once formula (2) is determined, the flattening rate of the three-dimensional image of the same lung to be tested is obtained using formula (2), and then the volume is estimated using formula (1). The voice sensor is used to capture sound signals for chest breathing sound recognition, environmental sound monitoring, etc. The voice sensor can obtain the chest sound of the patient when breathing, and then judge the condition of COPD; the data collected by the optical heart rate sensor is used to analyze the heart rate; the breathing sensor is used to obtain the breathing rate and breathing rhythm of COPD patients; the blood oxygen sensor and air quality sensor can obtain the patient's blood oxygen saturation changes and indoor air pollutant data.

9. The method for intelligent dynamic monitoring of COPD based on wireless sensing according to claim 7 is characterized by: In S2, for the i-th channel, the input noisy data X i ={X i，1 , X i，2 , ..., X i，N }, the process through the encoder is: Among them, θ i is the learning parameter of the encoder, N is the total number of samples in the i-th channel, h i is the potential representation of the sample; The potential representation is fed into the decoder, and the process through the decoder is in, is the learning parameter of the decoder, To reconstruct samples; the mean square error function is used during training in, is the regularization factor; the reconstructed sample obtained This is the denoised data.

10. The method for intelligent dynamic monitoring of COPD based on wireless sensing according to claim 7, characterized in that: In S3, first, the principal component analysis method is used to represent the multimodal data of m types and n samples using the matrix X as follows: Next, data standardization Among them, u is the average value, S j is the standard deviation; Calculate the standardized data correlation coefficient matrix R: where r ij is the correlation coefficient between the i-th row and j-th column of the standardized matrix Z; Solve for the eigenvalues ​​of the matrix R; use the Jacobian method | R-βI p |=0 to find the eigenvalue β of the correlation coefficient matrix R i (, sort the eigenvalues ​​from large to small; I p is the identity matrix; Calculate the proportion of each eigenvalue to the total number of eigenvalues, that is, the variance contribution rate: in, For the cumulative variance contribution rate, select the component F corresponding to the eigenvalue with the larger variance contribution rate C As the cth principal component; reconstruct the multimodal data after removing redundant information and reducing the dimension by c principal components, and mine the principal components related to audio modality, text modality and COPD; then, according to the data type and source, divide the original data into several independent channels, including breathing sound channel, chest volume change channel, and heart rate channel; the data of each channel is input into the denoising modality autoencoder respectively, and the feature extraction ability of the autoencoder is used to extract the robust multimodal features of each channel; The number of unlabeled samples input to the autoencoder is M, the feature dimension is D, the number of hidden layers of the autoencoder is L, and the number of neurons in the top layer of the hidden layer is S L , the weight of the autoencoder unit’s contribution to the final output features: The bias unit vector of each hidden layer constitutes the bias unit matrix of the model: Input sample vector e=1,2,…,M,the nonlinear representation in the lth hidden layer of the autoencoder is And the input layer l = 0, then The decoding process is actually the inverse process of encoding. The weight coefficient of the neuron propagation path is symmetrical with the encoding process. The decoding reconstruction of the output layer l = 2L is calculated, where L is the number of top layers of the hidden layer: The two important encoding parameter vectors of the communication radiation source signal sample in the autoencoder are recorded as a training parameter matrix: W=[(ω (1) ,b (1) ),(ω (2) ,b (2) )…,(oh (l) ,b (l) )] T (17) First, construct the objective function as: Among them, the reconstruction matrix Y is the self-encoder according to the encoding coefficient ω (l) and bias unit b (l) For the input sample matrix X U Reconstruction; γ is the weight attenuation coefficient to avoid overfitting of the objective function to the sample; s l Indicates that the current autoencoder is the sth l Layer; weight coefficient vector is the weight coefficient of the propagation path between the jth neuron in the lth layer and the ith neuron in the l+1th layer; Use the iterative optimal value method to solve the training parameter W that minimizes the objective function; The gradient of J(W) on the jth neuron in the lth layer of the network hidden layer is It constitutes the gradient matrix g(W)∈R SL×D ; Perform Taylor second-order expansion on J(W) (the subscript k+1 indicates the number of iterations): Derivation g(W)=g(W k+1 )+H k+1 (W-W k+1 ) (20) Among them, H k+1 is the k+1th calculation result of the Hessian matrix of the objective function; generally, H0 required for the first iteration is set to be the unit diagonal matrix, and the kth iteration value of W is recorded as W k , and continue to transform into In the above formula, let the training parameter change matrix s k =W k+1 -W k (22) Gradient change matrix y k =g k+1 -g k (23) B k+1 Represents the inverse of the Hessian matrix of the objective function to be estimated but B k+1 ·y k =s k (24) Let the auxiliary operator In the above process, s k With y k All depend on the training parameters required W pre =[ω pre ,b pre ] (26) ω pre is the weight of each hidden layer of the autoencoder contributing to the final output feature, b pre is the bias unit vector of each hidden layer; Each B in the k+1 iteration formula k+1 , solve the sample feature W pre ∈R D×N , the training parameters extracted at this time are the desired features; The multimodal feature outputs extracted by the denoising modal autoencoder are spliced ​​in the channel dimension, and the channel attention mechanism is adopted. Through compression and excitation, the network pays attention to the mutual influence relationship between different channels, learns the importance of each channel feature, extracts the key information in the feature, and performs dimensional transformation of the fully connected layer to achieve feature fusion; the compression and excitation network of the channel attention mechanism is mainly composed of compression units and excitation units; the compression unit uses the pooling layer to sum and compress the global spatial features in each channel to form their own channel features; pooling divides the feature space into multiple subspaces, selects a value in each subspace, and forms a new feature space with the selected values ​​to reduce the dimension of the feature space; the excitation unit uses two fully connected layers to obtain the weight parameters w1 and w2 of each layer to reduce the number of parameters of the unit and enhance the migration ability of the unit; during the training process of the model, the feature weights of each channel domain and the correlation between channels are obtained; the calculation formula of the channel attention mechanism is as follows: you c =σ(w2δ(w1u s )) (28) Among them, σ is the summation function; δ represents the δ function, also known as the impulse function; Scale is the dimensionality transformation; u s is the normalized global information of the concatenated output, T is the time step, u c is the weight corresponding to each channel, z(i) represents the output result after splicing, It is the weighted output; finally, a dimensional transformation of the fully connected layer is performed to achieve feature fusion.

Images