CN108665066A - Portable respirator flow curve scaling method based on BP neural network - Google Patents
Portable respirator flow curve scaling method based on BP neural network Download PDFInfo
- Publication number
- CN108665066A CN108665066A CN201810469621.2A CN201810469621A CN108665066A CN 108665066 A CN108665066 A CN 108665066A CN 201810469621 A CN201810469621 A CN 201810469621A CN 108665066 A CN108665066 A CN 108665066A
- Authority
- CN
- China
- Prior art keywords
- neural network
- portable respirator
- flow curve
- iterations
- trained
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Measurement Of The Respiration, Hearing Ability, Form, And Blood Characteristics Of Living Organisms (AREA)
Abstract
The present invention relates to a kind of portable respirator flow curve scaling method based on BP neural network, includes the following steps:Step 1:Calibration experiment is carried out using portable respirator, the carbon dioxide flow data acquired are divided into two groups, one group is trained, and one group is tested;Step 2:Establish BP neural network, S function is chosen as excitation function, it is trained using LM algorithms, the neural network of use is made of for two layers hidden layer and output layer, and hidden layer includes 20 neurons, with the increase of iterations, least mean-square error reduces, less than iterations be 6 when, stop study, finally obtain flow curve.
Description
Technical field
It is to a kind of new of people's End-tidal carbon dioxide flow data collector the invention belongs to biomedical engineering field
Fit approach.
Background technology
Breathing is a kind of important metabolic processes, it is the sign of life for maintaining people to stablize.Respiratory tract disease at present
Disease, such as asthma, bronchiectasis and obstructive sleep apnea, are widely current in the whole world.According to 2016 world healths
The investigation of tissue suffers from asthma more than 200,000,000 3,500 ten thousand people, has more than 3,000,000 people every year and die of chronic obstructive pulmonary disease (COPD).
In order to which effectively, accurately evaluation cardio-pulmonary function, monitoring of respiration are an important compositions in respiratory disease and ICU management
Part.
At present during biomedical respiratory flow measures, the fitting of data is highly important, is carried out to data
Effective fitting can better response measurement waveform, increase the precision of measurement.In previous measurement process, often make
The fit approach of sensor is least square method, the fit approach such as difference fitting, but during measurement, especially
Sensor accuracy be not it is very high in the case of these approximating methods will produce larger error.
Current portable respirator effectively can add the differential pressure of acquisition by sensor when measuring
Work converts, but during conversion, since the sensor and processing circuit of portable respirator all cannot be big with profession
Type lung ventilator is compared, so a kind of fitting algorithm of profession is essential.
Invention content
The object of the present invention is to provide a kind of portable respirator flow curve scaling methods, to improve sensor accuracy.
Technical solution is as follows:
The error that pipeline and sensor are brought when measuring can be effectively improved when measuring, and obtained fine
Fitting effect, operating process is:
A kind of portable respirator flow curve scaling method based on BP neural network, includes the following steps:
Step 1:Calibration experiment is carried out using portable respirator, the carbon dioxide flow data acquired are divided into two
Group, one group is trained, and one group is tested;
Step 2:BP neural network is established, S function is chosen as excitation function, is trained using LM algorithms, use
Neural network is made of for two layers hidden layer and output layer, and hidden layer includes 20 neurons, with the increase of iterations, it is minimum
Square error reduces, and when to be less than iterations be 6, stop study, finally obtains flow curve.
The present invention overcomes the problems, such as that the fitting precision that common fitting algorithm is brought is low, effectively improves portable respirator
Measurement accuracy.
Description of the drawings
Fig. 1 Tan-sigmoid functions
Fig. 2 prediction data and real data
Fig. 3 neural network parameters
Fig. 4 neural network fitting results are analyzed
The percentage error of Fig. 5 BP neural networks fitting
Specific implementation mode
The present invention is optimized according to carbon dioxide concentration measurement device, and design one kind can effectively improve titanium dioxide
The fitting algorithm of carbon measurement accuracy.
In view of mainstream equipment must have the characteristics that stablize and real-time, so setting of being combined of pneumatic duct respiratory circuit
Meter has to comply with this requirement.Airway pressure is monitored by sampling pressure, and airway pressure has close pass with air flue
System.The main consideration of pneumatic duct design is to will produce pressure drop after air flows through air flue.The design of gas pipeline is to be based on Bernoulli Jacob
Law and continuity law, by continuity equation ρ1υ1s1=ρ2υ2s2=m is applied in flow monitoring:
ρsAvA=ρ υBsB=m formulas (1)
Wherein it is sAAnd sBThe area in section, ρ are the density of air, vAAnd vBIndicate the speed of the air-flow of two ports, m
The quality of gas.
In conjunction with formula (2)
Two formula, which are merged, can find out the relationship of pressure difference and flow:
Wherein Q is volume flow.
In order to accurately be fitted the function curve of voltage and flow, it is very to select suitable excitation function and training function
Important.The present invention chooses S function as excitation function, is trained using Levenberg-Marquardt (LM) algorithm.
Excitation function can be defined as:
The range of output function is from -1 to 1, such as Fig. 1.
It is a kind of using mark that we, which have selected algorithms of the Levenberg-Marquardt (LM) as training data, LM algorithms,
The fast algorithm of quasi- numerical optimization, it is the combination of gradient descent method and Gauss-Newton methods.LM algorithms not only have
The local convergence of Gauss-Newton methods, and it is of overall importance with gradient descent method.
LM algorithms utilize second dervative information, much faster than gradient descent method.The vector formulas of weight and threshold value is:
X=[wih(1,1)...wih(h,i),bn(1)...bn(h)wh0(1,1)...b0(0)]TFormula (5)
Therefore, the vector being made of update weights and threshold value is:
X (k+1)=x (k)+Δ x formulas (6)
Wherein Δ x represents the change of weights and threshold value, and LM algorithms are improved Gauss-Newton algorithms, can be defined
For:
Δ x=- [JT(x)J(x)+μI]-1JT(x) e formulas (7)
Wherein J (x) is Jacobin matrix, and μ is damped coefficient, and I is unit matrix.
LM algorithms are similar to gradient descent method.Each iteration reduces the value of figure layer.So when algorithm is close to target error,
It moves closer in GaussNewton algorithms.
LM algorithms are a kind of effective algorithms, and the basic thought of iterative process allows the progress mistake on degeneration direction to search
Rope.Meanwhile in order to enable the weights and threshold value of neural network all achieve the effect that it is satisfied, we open gradient descent algorithm and
The algorithm of adaptive neural network improves extensive ability.
The error that pipeline and sensor are brought when measuring can be effectively improved when measuring, and obtained fine
Fitting effect, operating process is:
Step 1:Calibration experiment is carried out using portable respirator
Step 2:60 groups of data are sampled, changes in flow rate is at the uniform velocity promoted from 1L/min to 60L/min, is divided into 1L/
Min acquires corresponding voltage value.Data flow is divided into two groups, and the data on flows of odd number group is trained, and obtains even number set flow
Data are tested.
After training, test data is inputted into neural network.The shape and computational accuracy index of matched curve are described.It should
Neural network used by scaling method is made of for two layers hidden layer and output layer.Hidden layer includes 20 neurons.Iterations 6
It is secondary, time 1s.When iteration reaches 6 times, Grad 0.00496.Fig. 2 indicates the training and survey of fitting degree between the two
Try data and reality output data.
With the increase of iterations, least mean-square error reduces, when to be less than iterations be 6.Then it stops
Study, will eventually get specific curve.
Step 3:Trained m odel validity is analyzed such as Fig. 3
Fig. 3 shows the learning process under network condition, with the increase of iterations, it may be seen that LM algorithms
Graded.Mu is the average value of normal distribution, similar to error.With the increase of iterations, if error amount increases, Mu
Value also can be with increase, when the value of Mu increases to a certain size, learning process will stop.
Step 4:The precision that intuitive analysis intuitively shows fitting, Fig. 4 are carried out to the fitting result of data in the measurements
The experimental data that obtained output data and actual experiment obtain after BP neural network training is set forth.We can be direct
Observe error.The distance between most of output datas and reality output data very little, only a small number of data generate larger
Error.The trend of curve matching is unaffected, and plots changes reflect the relationship of outflow and voltage.Therefore, we
It can conclude that:The accuracy of the training of neural network is within error range.
Step 5:Quantitative analysis is carried out to error, from Fig. 5 it will be seen that in BP neural network, training data exists
Individual data items are there are larger error, but the fluctuation very little of training error.Finally we obtain absolute error 0.7594.This
Number is very intuitive, it is shown that the overall precision of neural metwork training.Fig. 5 gives at the BP neural network of 30 groups of test datas
Test exports the percentage of the error between reality output data after reason.Mathematical formulae is defined as:
Wherein m is test output data, and n is reality output data.They are in 4L/min, 12L/min as can be known from Fig. 5
Larger error is generated with 40L/min, but it only occurs in seldom data, the nerve net without influencing good fitting function
Network.
Claims (1)
1. a kind of portable respirator flow curve scaling method based on BP neural network, includes the following steps:
Step 1:Calibration experiment is carried out using portable respirator, the carbon dioxide flow data acquired are divided into two groups, one
Group is trained, and one group is tested;
Step 2:BP neural network is established, S function is chosen as excitation function, is trained using LM algorithms, the nerve of use
Network is made of for two layers hidden layer and output layer, and hidden layer includes 20 neurons, and with the increase of iterations, lowest mean square misses
Subtractive is small, when being less than iterations position 6, stops study, finally obtains flow curve.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810469621.2A CN108665066A (en) | 2018-05-16 | 2018-05-16 | Portable respirator flow curve scaling method based on BP neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810469621.2A CN108665066A (en) | 2018-05-16 | 2018-05-16 | Portable respirator flow curve scaling method based on BP neural network |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108665066A true CN108665066A (en) | 2018-10-16 |
Family
ID=63778385
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810469621.2A Pending CN108665066A (en) | 2018-05-16 | 2018-05-16 | Portable respirator flow curve scaling method based on BP neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108665066A (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111383764A (en) * | 2020-02-25 | 2020-07-07 | 山东师范大学 | Correlation detection system for mechanical ventilation driving pressure and related events of breathing machine |
CN115970118A (en) * | 2022-12-22 | 2023-04-18 | 宁波市计量测试研究院(宁波新材料检验检测中心) | Calibrating device for respiratory humidification therapeutic instrument |
-
2018
- 2018-05-16 CN CN201810469621.2A patent/CN108665066A/en active Pending
Non-Patent Citations (1)
Title |
---|
DAYONG FAN ET.AL.: ""Effectively Measuring Respiratory Flow With Portable Pressure Data Using Back Propagation Neural Network"", 《IEEE JOURNAL OF TRANSLATIONAL ENGINEERING IN HEALTH AND MEDICINE》 * |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111383764A (en) * | 2020-02-25 | 2020-07-07 | 山东师范大学 | Correlation detection system for mechanical ventilation driving pressure and related events of breathing machine |
CN111383764B (en) * | 2020-02-25 | 2024-03-26 | 山东师范大学 | Correlation detection system for mechanical ventilation driving pressure and ventilator related event |
CN115970118A (en) * | 2022-12-22 | 2023-04-18 | 宁波市计量测试研究院(宁波新材料检验检测中心) | Calibrating device for respiratory humidification therapeutic instrument |
CN115970118B (en) * | 2022-12-22 | 2023-08-04 | 宁波市计量测试研究院(宁波新材料检验检测中心) | Respiratory humidification therapeutic instrument calibrating device |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Rissler et al. | Deposition efficiency of inhaled particles (15-5000 nm) related to breathing pattern and lung function: an experimental study in healthy children and adults | |
Choi et al. | Mathematical analysis of particle deposition in human lungs: an improved single path transport model | |
CN102770069B (en) | Nitric oxide measurement method and apparatus | |
Mori et al. | Colored 3-dimensional analyses of respiratory resistance and reactance in COPD and asthma | |
Archer et al. | Comparing aerosol number and mass exhalation rates from children and adults during breathing, speaking and singing | |
Gillespie et al. | Aerodynamic profiles of women with muscle tension dysphonia/aphonia | |
CN102469954A (en) | Method and apparatus of determining exhaled nitric oxide | |
CN108109697A (en) | A kind of pulmonary function test (pft) system and method based on expiration mathematical model | |
CN108665066A (en) | Portable respirator flow curve scaling method based on BP neural network | |
CN105078462B (en) | The method and apparatus for estimating vapour lock and compliance | |
SE1230100A1 (en) | System for optimal mechanical ventilation | |
Füri et al. | Comparison of airway deposition distributions of particles in healthy and diseased workers in an Egyptian industrial site | |
US10188317B2 (en) | Measuring respiratory mechanics parameters using perturbations | |
Sun et al. | C-RIDGE: Indoor CO2 data collection system for large venues based on prior knowledge | |
CN106644361B (en) | A kind of simple and easy method measuring transonic wind tunnel test section space flow field symmetry | |
CN206239414U (en) | For the flow sensor of pulmonary function detection | |
Kazemi et al. | Frequency analysis of capnogram signals to differentiate asthmatic and non-asthmatic conditions using radial basis function neural networks | |
Si et al. | Deciphering exhaled aerosol fingerprints for early diagnosis and personalized therapeutics of obstructive respiratory diseases in small airways | |
Millage et al. | A review of inhalability fraction models: discussion and recommendations | |
CN109659031A (en) | Lung function index prediction meanss and determining method | |
Ionescu et al. | A recurrent parameter model to characterize the high-frequency range of respiratory impedance in healthy subjects | |
CN107095676A (en) | The method of testing and system of lung function instrument data processing accuracy | |
Bekdash et al. | Standard testing procedure for quantifying breathing gas carbon dioxide partial pressure for extravehicular activity and launch, entry, survival pressure suits | |
Invernizzi et al. | Real-time measurement of particulate matter deposition in the lung | |
Curran et al. | Using acoustic sensors to discriminate between nasal and mouth breathing |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20181016 |
|
RJ01 | Rejection of invention patent application after publication |