CN106249159B - A kind of electric quantity monitoring method and electric quantity monitoring system of brain pacemaker - Google Patents
A kind of electric quantity monitoring method and electric quantity monitoring system of brain pacemaker Download PDFInfo
- Publication number
- CN106249159B CN106249159B CN201610629704.4A CN201610629704A CN106249159B CN 106249159 B CN106249159 B CN 106249159B CN 201610629704 A CN201610629704 A CN 201610629704A CN 106249159 B CN106249159 B CN 106249159B
- Authority
- CN
- China
- Prior art keywords
- module
- data
- electric quantity
- brain pacemaker
- quantity monitoring
- 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.)
- Active
Links
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R31/00—Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
- G01R31/36—Arrangements for testing, measuring or monitoring the electrical condition of accumulators or electric batteries, e.g. capacity or state of charge [SoC]
- G01R31/367—Software therefor, e.g. for battery testing using modelling or look-up tables
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R31/00—Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
- G01R31/36—Arrangements for testing, measuring or monitoring the electrical condition of accumulators or electric batteries, e.g. capacity or state of charge [SoC]
- G01R31/382—Arrangements for monitoring battery or accumulator variables, e.g. SoC
Abstract
The present invention relates to a kind of electric quantity monitoring method of brain pacemaker and using the electric quantity monitoring system of this method.This method comprises: obtaining the stimulation parameter of brain pacemaker;The battery relevant parameter of the brain pacemaker is calculated by SVM prediction model;And the battery parameter being calculated is shown to user.The electric quantity monitoring method and electric quantity monitoring system of brain pacemaker provided by the invention may be implemented to carry out high-precision prediction to the electricity to brain pacemaker, and the prediction scheme may be implemented to combine and the DBS electric quantity monitoring and life appraisal of individual patients situation with what clinical specificity combined for each stimulation parameter, have certain reliability, finally can be applied to clinic.
Description
Technical field
The present invention relates to medical instrument correlative technology fields, in particular, being related to a kind of implantable medical devices
(Implantable Medical Device,IMD)。
Background technique
Brain pacemaker, also known as lesions located in deep brain (Deep Brain Stimulation, DBS) are current treatment advanced stages
With the effective technology of drug refractory dyskinesia and mental disorder.Curative effect is aobvious especially in the treatment of Parkinson's disease
It writes.Recent research achievement shows that brain pacemaker can also be used for the treatment mentals disease such as Alzheimer's disease and obsessive-compulsive disorder, depression
Disease.Referring to Fig. 1, the implementation device of lesions located in deep brain is brain pacemaker, it is a set of implantable microelectronic device, including
Impulse generator (IPG), electrode and extension wire.After brain pacemaker barbing swashs, impulse generator will be set according to vitro program controlled instrument
Fixed stimulation parameter, which generates, continues electric pulse, and electric current is stimulated in subthalamic nuclei or globus pallidus via extension wire by electrode contacts
The nerve nucleus of side core.Since these components are both needed to implant, brain pacemaker has low in energy consumption, small in size, safety
The features such as strong, while having higher requirement to its material property, temperature characterisitic and configuration design.It can by vitro program controlled instrument
To adjust all stimulation parameters including stimulation amplitude, frequency and pulsewidth.The battery that brain pacemaker uses includes rechargeable battery
And non-rechargeable battery.
Clinically need to know charge level and the residue longevity of the non-rechargeable battery of the DBS of brain pacemaker implantation person
Life.The battery life predicting of inaccuracy, which will will lead to, prematurely carries out DBS replacement operation, causes waste and the patient's property of battery
Loss;Or carry out replacement operation too late and cause hang-up, it is unfavorable to therapeutic effect.For DBS eligible patients, art
It is the means for guaranteeing that curative effect is mostly important that rear stimulation parameter (amplitude, frequency, pulsewidth), which is adjusted, and almost every DBS implantation person is
The process adjusted and joined can be undergone, current life prediction scheme is only applicable to typical (common) stimulation parameter combination, and stimulates ginseng
Several adjusts the accumulation that will all bring error each time, this is very unreliable prediction result so that can not be applied to clinic.But
From the point of view of reality, the method that this approximate evaluation can only be used at present, because under the premise of not considering impedance, amplitude,
As soon as the combination of three frequency, pulsewidth stimulation parameters there are more than 100 ten thousand kinds, wherein there are more than 50 ten thousand kinds of combinations under voltage mode, it will be every
Power consumption under kind of parameter combination all measure come it is clearly unpractical.
Therefore, it is necessary to seek high-precision prediction technique to carry out power quantity predicting to brain pacemaker, and require the prediction side
Case may be implemented to combine and (such as impedance) of individual patients situation with what clinical specificity combined for each stimulation parameter
DBS electric quantity monitoring and life appraisal, and can provide corresponding error range and characteristic distributions, have certain precision and can
By property, clinic finally can be applied to.
Summary of the invention
The present invention provides the electric quantity monitoring method and electric quantity monitoring system of a kind of brain pacemaker.
A kind of electric quantity monitoring method of brain pacemaker, this method comprises: obtaining the stimulation parameter of brain pacemaker;Pass through support
Vector machine forecast model calculates the battery relevant parameter of the brain pacemaker;And the battery parameter being calculated is shown to use
Family.
According to the electric quantity monitoring method of above-mentioned brain pacemaker, wherein the SVM prediction model passes through with lower section
Method is established: step S10 determines selected kernel function, enters step S11;Step S11 initializes the parameter of kernel function, into step
Rapid S12;Step S12 obtains experiment sample data, enters step S13;Sample data is used to learn or predict pair by step S13
Than if it is study, entering step S14, being compared if it is prediction, then enter step S18;Step S14 carries out sample data
Pretreatment, enters step S15;Step S15 judges whether selected parameter attribute is representative, if so, S16 is entered step,
If it is not, then return step S14;Step S16, algorithm training establish model, enter step S17;Step S17, is searched using grid
Suo Fangfa Optimal Parameters, enter step S18;Step S18, prediction data and experimental data are compared, and enter step S19;
Step S19, error in judgement whether within the allowable range, if so, S20 is entered step, if it is not, then return step S16;And
Step S20, as prediction model.
According to the electric quantity monitoring method of above-mentioned brain pacemaker, wherein in the step S10, selected kernel function are as follows: radial
Base kernel function K (xi,xj)=exp (- γ | | xi-xj||2),γ>0。
According to the electric quantity monitoring method of above-mentioned brain pacemaker, wherein in the step S12, obtain experiment sample data
Method are as follows: the data of three dimensions, amplitude, frequency and pulsewidth are acquired using Auto-Test System.
According to the electric quantity monitoring method of above-mentioned brain pacemaker, wherein in the step S14, located in advance to sample data
The method of reason is to normalize or take sequence number.
According to the electric quantity monitoring method of above-mentioned brain pacemaker, wherein the normalization formula (1) are as follows: X1=2 ((X-
Xmin)/(Xmax-Xmin)) -1 (1), wherein X is former data;X1The value for being X after normalized;XmaxAnd XminIt is respectively former
Maximum value and minimum value where data X in data group.
According to the electric quantity monitoring method of above-mentioned brain pacemaker, wherein the method for taking sequence number are as follows: amplitude sequence number=
Amplitude × 20;Pulsewidth sequence number=(pulsewidth -30)/10;Frequency sequence number is numbered according to natural number with 1 for increment.
A kind of brain pacemaker electric quantity monitoring system comprising: a control module, the control module control entire brain pace-making
The work of device electric quantity monitoring system;One computing module, the computing module are used to calculate the brain pacemaker by the above method
Battery relevant parameter;One display module, the display module be used for by the computing module be calculated as the result is shown to use
Family;One data input module, the data input module are used to input the data information of brain pacemaker;And a memory module,
Memory module information for storing data.
It further comprise a data acquisition module and a communication module according to above-mentioned brain pacemaker electric quantity monitoring system;Institute
Data acquisition module is stated for being communicated by the vitro program controlled instrument of the communication module and the brain pacemaker, to obtain brain
The stimulation parameter of pacemaker.
It further comprise a comparison module and a cue module according to above-mentioned brain pacemaker electric quantity monitoring system;The ratio
It is used to for the result that the computing module is calculated being compared with a secure threshold compared with module, thus in danger by being somebody's turn to do
Cue module is prompted;And the result that the computing module is calculated by a graphic user interface for the cue module
It is shown to user.
Compared to the prior art, the electric quantity monitoring method of brain pacemaker provided by the invention and electric quantity monitoring system can be with
It realizes and high-precision prediction is carried out to the electricity to brain pacemaker, and the prediction scheme may be implemented mutually to tie with clinical specificity
The DBS electric quantity monitoring and life appraisal for each stimulation parameter combination and individual patients situation closed, have it is certain can
By property, clinic finally can be applied to.
Detailed description of the invention
Fig. 1 is the working principle diagram of brain pacemaker used in the embodiment of the present invention.
Fig. 2 is structural risk minimization schematic diagram.
Fig. 3 is the feature space schematic diagram of support vector machines.
Fig. 4 is kernel function schematic diagram.
Fig. 5 is SVM prediction model modeling procedure figure.
Fig. 6 is that SVM does not carry out pretreated prediction absolute error to trained and test sample and Curve Fitting Prediction is absolute
Error comparison diagram.
Fig. 7 is SVM not opposite with Curve Fitting Prediction to the pretreated Relative Error of trained and test sample progress
Error comparison diagram.
Fig. 8 is that SVM carries out taking the pretreated prediction absolute error of sequence number and curve matching pre- to trained and forecast sample
Survey absolute error comparison diagram.
Fig. 9 is that SVM carries out taking the pretreated Relative Error of sequence number and curve matching pre- to trained and forecast sample
Survey relative error comparison diagram.
Figure 10 is SVM samples normalization and the prediction absolute error comparison diagram for taking sequence number processing.
Figure 11 is SVM samples normalization and takes sequence number Relative Error comparison diagram.
Figure 12 is the preferred mse isogram of SVM parameter.
Figure 13 is the structural schematic diagram of brain pacemaker electric quantity monitoring system provided by the invention.
Figure 14 is the graphic user interface schematic diagram of brain pacemaker electric quantity monitoring system provided by the invention.
Main element symbol description
Brain pacemaker electric quantity monitoring system | 10 |
Control module | 100 |
Computing module | 101 |
Data acquisition module | 102 |
Communication module | 103 |
Display module | 104 |
Data input module | 105 |
Memory module | 106 |
Comparison module | 107 |
Cue module | 108 |
Following specific embodiment will further illustrate the present invention in conjunction with above-mentioned attached drawing.
Specific embodiment
The present invention provides a kind of electric quantity monitoring method of brain pacemaker and using the brain pacemaker electricity prison of this method
Examining system.This method can predict the electricity of non-rechargeable battery and the electric quantity consumption situation of rechargeable battery.To not chargeable
Battery is particularly important.
The present invention introduces support vector machines (SVM) prediction model first.
Statistical Learning Theory (Statistic Learning Theory, SLT) is that one kind is specialized under Small Sample Size
The basic theories and mathematical construct of machine learning rule.The Learning machine branch based on structural risk minimization proposed by Vapnik
It holds vector machine (Support Vector Machine, SVM) and is used as a kind of very potential recurrence sorting technique, be a kind of base
In the mode identification method of Statistical Learning Theory.Function regression estimation is a kind of common Machine Learning Problems, and SVM is asked at this
By the complexity and approximation accuracy of control function simultaneously in topic, good Generalization Ability is obtained, with the side of SVM approximating function
Method is support vector regression (Support Vector Regression, SVR).
SVM method is built upon on the basis of VC dimension theory and the structural risk minimization principle of Statistical Learning Theory, root
Seek best compromise between the complexity and learning ability of model according to limited sample information, is that one kind specializes in small sample
In the case of machine learning rule theory.Statistical inference rule under this theory not only allows for wanting progressive performance
It asks, and pursues and obtain optimal result under conditions of existing limited information.Other learning methods such as SVM theory and neural network
It compares, the parameter with core can be calculated automatically by the method for optimization, and avoid local minimum points, overfitting
The defects of.In summary, SVM method is not only able to well solve for regression estimation problem perplexed many learning methods in the past
The practical challenges such as small sample, overfitting, high dimension, local optimum, and have very strong generalization ability.
On the one hand, the precision of the least square regression model power consumption prediction based on circuit theory is limited, and only in allusion quotation
The combination of type stimulation parameter and impedance nearby have the preferable precision of prediction relative to own.And power consumption is pre- under SARS shape parameter
It is then very low to survey precision.On the other hand, since under the premise of not considering that different impedances and battery power voltage influence, light is in electricity
Stimulation parameter combination just has more than 50 ten thousand kinds under die pressing type, if again taken into account impedance, this numerical value will be with multiple
Form increases.And the power consumption number of every one parameter combination of acquisition of Auto-Test System needs at least half a minute, by the irritating ginseng of institute
Power consumption under array conjunction is all acquired to come out and is difficult to realize.In view of the irritating amplitude of the problem, frequency, pulsewidth and impedance four variations
Stimulation parameter, belong to higher-dimension, small sample (can only at most acquire more than 20,000 parameter combinations), nonlinear problem, be difficult directly into
The fitting of row data.It thus attempts to be predicted using support vector machines, and obtains good prediction effect, power consumption can be made pre-
Survey error and be much smaller than least square regression model, while its error is more evenly distributed, is more flat, predict power consumption absolute error and
Relative error trend is complementary, is more advantageous to the electric quantity monitoring and life appraisal for carrying out the later period, and the unlikely long-term accumulation because of error is given
Life prediction brings too much influence.
Basic thought and theoretical basis of the SVM for regression forecasting is described below.
Neural network is to be based on empirical risk minimization principle, but really be able to directly affect model forecast result
Be expected risk practical risk in other words, and structure risk directly determines expected risk.Empirical risk minimization is difficult to ensure
Practical risk is minimum, this is that neural network the reason of over-fitting easily occurs.Before explaining over-fitting, it is appreciated that trained number first
It is two different concepts according to collection precision and test data set precision.So-called over-fitting, it is simple understand be exactly, if one point
Class device, it is higher in training dataset precision, it is lower in test data set precision, just illustrate over-fitting occurred.
Different from neural network, support vector machines are to be based on structural risk minimization, and structure risk is by passing through
Test what risk and fiducial range codetermined.Wherein empiric risk characterizes the fitting precision of model, and fiducial range then represents mould
The Generalization Ability of type.Referring to fig. 2, if thinking expected risk minimum, empiric risk and fiducial range is needed to reach a balance,
It is exactly that can make Structural risk minization at point h*, has both been not in owe study overfitting phenomenon occur, so that model
Fitting precision and precision of prediction it is all higher.In short, SVM based in structural risk minimization include two optimizations refer to
Mark: empiric risk and confidence interval.The two has codetermined the practical risk of supporting vector.For linear separability data, SVM's
Principle is fixed empiric risk, optimizes VC fiducial range.This just needs to minimize VC dimension, has finally swung to most basic ask
Topic, exactly maximizes Margin.From this, which becomes pure mathematics problem.And for linearly inseparable data, then it needs to add
Enter penalty factor or improve dimension, is transformed into linear separability problem.In contrast, the machine learning methods such as neural network
All it is fixed fiducial range, minimizes empiric risk, therefore easily occur that fitting precision is very high and precision of prediction is very low, that is, mistake
The case where fitting.SVM is based on structural risk minimization, can maximize fiducial range while guaranteeing fitting precision to guarantee
Its Generalization Ability.So the generalization ability of SVM is better than neural network.
Referring to Fig. 3, the basic thought of support vector machines is the concept in introduced feature space, passes through a Nonlinear Mapping handle
Nonlinear problem on former number field is converted into the linear problem on feature space, while also introducing the concept of kernel function, feature
Linear problem spatially, which is restored to again on former number field, to be implemented, without regard to the concrete form of feature space and Nonlinear Mapping,
Best Generalization Ability can be thus obtained while not improving computation complexity.
In order to construct optimal separating hyper plane in high-dimensional feature space, do not need to consider feature sky with Explicit Form
Between, and only it is required to calculate the inner product of vector in supporting vector and feature space, that is, calculate in a manner of kernel function.
It is had the advantage that, can be converted in the Nonlinear Classification face in the input space in the feature space F of higher-dimension using kernel function
Linear hyperplane handle.Referring to fig. 4, it is kernel function schematic diagram, embodies and be mapped to the Nonlinear Classification of the input space
The process of the linear classification of feature space.
In summary, sample x can be mapped to high-dimensional feature space F by Nonlinear Mapping ψ (x) using SVM, and
Optimum regression function is solved in F.Briefly, support vector machines may be implemented to convert height for the nonlinear regression of lower dimensional space
Linear regression in dimension space.And kernel function is then by way of calculating supporting vector and inner product of vectors in feature space by higher-dimension
Linear problem in space is restored to the former input space again to be implemented.Therefore, kernel function K appropriate is used in optimum regression function
(x, xi) replaces the inner product of vectors ψ (x in higher dimensional spacei) ψ (x), so that it may the Linear Quasi after realizing a certain nonlinear transformation
It closes, and computation complexity does not increase.It may be said that kernel function plays important angle for algorithm of support vector machine
Color.Therefore the selection of kernel function is extremely important, will make a concrete analysis of below.
The heterogeneous linear recurrence of support vector regression and nonlinear regression.For linear regression, consider to use linear regression function (1)
Y=f (x)=ω x+b (1)
For nonlinear regression, then sample x is mapped to high-dimensional feature space F by Nonlinear Mapping ψ (x), and in F
Optimum regression function is solved, and the inner product of vectors ψ (x in higher dimensional space is replaced using kernel function K (x, xi) appropriatei)·ψ
(x), to realize the linear fit after a certain nonlinear transformation.
F (x, ω)=(ω ψ (x))+b (2)
Wherein, ψ (x) is the nonlinear transformation that sample point is mapped to higher dimensional space.
SVM regression machine is represented by
Wherein, ξiWithFor slack variable,The first two constraint condition that formula (3) cannot be fully met (do not include
ξiWith) under introduce.
Meet constraint:
Majorized function (3) is quadratic form, and constraint condition is linear, therefore is a typical quadratic programming problem, can be used
Method of Lagrange multipliers solves.Introduce Lagrange multiplierIt can obtain
Have at optimal solution
Formula (6) are substituted into formula (5), the former problem under the conditions of linear separability is transformed to its dual problem, is solved following
Maximization problem:
The Lagrange dual problem for solving above-mentioned model, obtains regression function:
Wherein, K (xi, x) and=ψ (xi) ψ (x) be kernel function.
There are three very important parameters in this model, are penalty coefficient C respectively, the error range ε of permission is (in mould
When type training ε be selected as 0.01), nuclear parameter γ.Penalty coefficient C is the equal of the tolerance to error, and C is bigger to illustrate that you more cannot
There is error in tolerance;Nuclear parameter γ impliedly determines that data are mapped to the distribution of new feature space.Generally, C is bigger, intends
It is smaller to close error, prediction error is bigger;ε is bigger, and prediction error is bigger, without apparent dull increase between error of fitting and ε
Or reduced relationship.Specific rule no for the selection of these three parameters is general to be provided by testing.Due to that will allow
Error range ε is chosen to be 0.01, and the parameter for carrying out selection is also needed to have penalty factor and nuclear parameter γ, best group about C and γ
The selection of conjunction can be described in detail below.
SVM prediction model modeling method process introduced below.SVM modeling method is programmed by Matlab and is realized.Referring to figure
5, SVM prediction model modeling methods the following steps are included:
Step S10 determines selected kernel function, enters step S11;
Step S11 initializes the parameter of kernel function, enters step S12;
Step S12 obtains experiment sample data, enters step S13;
Sample data is used to learn or predict to compare, if it is study, S14 is entered step, if it is pre- by step S13
Comparison is surveyed, then enters step S18;
Step S14, pre-processes sample data, enters step S15;
Step S15 judges whether selected parameter attribute is representative, if so, S16 is entered step, if it is not, then returning
Return step S14;
Step S16, algorithm training establish model, enter step S17;
Step S17 enters step S18 using trellis search method Optimal Parameters;
Step S18, prediction data and experimental data are compared, and enter step S19;
Step S19, error in judgement whether within the allowable range, if so, S20 is entered step, if it is not, then return step
S16;And
Step S20, as prediction model.
In the above method, main three step is step S10, the selection of kernel function;Step S14, data prediction;And step
Rapid S17, parameter optimization.
In the step S10, the kernel function used in support vector machines mainly has four classes:
(1) linear kernel function K (xi, xj)=xi·xj
(2) Polynomial kernel function K (xi, xj)=[γ (xi·xj)+coef0]d, γ > 0
(3) Radial basis kernel function (Gauss, RBF) K (xi, xj)=exp (- γ | | xi-xj||2), γ > 0
(4) Sigmoid function K (xi, xj)=tanh (v (xi·xj)+coef0)
For algorithm of support vector machine, the selection of kernel function is very crucial.Typically, Radial basis kernel function is (high
This kernel function, RBF kernel function) it is reasonable first choice.Sample is non-linearly mapped to the sky of a more higher-dimension by this kernel function
Between, different from linear kernel, it is capable of handling the non-linear relation of classification annotation (input feature vector) and attribute (output power consumption number).And
And linear kernel is a special case of RBF, therefore, uses the linear kernel of a penalty factor and the RBF of certain parameters (C, γ)
Core performance having the same.Meanwhile the performance of Sigmoid core is like the RBF core of certain parameter.Second reason, hyper parameter
(hyperparameter) quantity influences whether the complexity (because the selection of parameter can only be by test) of model selection, and more
Xiang Shihe ratio RBF core has more hyper parameters.Finally, RBF core has less numerical complexity (numerical
Difficulties), and polynomial kernel is related to high exponent arithmetic(al).In addition, Sigmoid core is not legal (example under certain parameters
Such as: not being the inner product of two vectors).It certainly, is not applicable there is also some situation RBF cores.Particularly, when intrinsic dimensionality is non-
When often big, it is likely that linear kernel can only be applicable in.Linear kernel is equivalently employed without, sample is mapped into higher dimensional space, linearly
It distinguishes (or recurrence) to be completed in original feature space, this is most fast selection.
The above analysis, due to either small sample or large sample, situations such as higher-dimension or low-dimensional, RBF kernel function
It is applicable in, there is wider convergence domain.And for nonlinear model, RBF kernel function is ideal recurrence foundation
Function, therefore select radial base RBF kernel function.
It is worth noting that, all comprising in adjustable parameter, such as Polynomial kernel function in the above kernel function
Parameter γ, coef0, d;The parameter v, coef0 etc. in parameter γ and Sigmoid kernel function in RBF kernel function.According to core letter
The difference that penalty factor and the middle error range ε allowed of formula (4) choose in number parameter and formula (3), the precision of prediction of model can
Very big difference can be had.Here it is the problem of parameter selection of SVM method, also known as problem of model selection.It will carry out below specific
Analysis.
In the step S12, the data acquired will be needed to be divided into three groups, totally 14561 samples:
First group: totally 630.Amplitude (V): 0-0.9 (step:0.1);Frequency (Hz): 2,30,70,110,130,150,
170,200,250;Pulsewidth (μ s): 30,70,110,150,200,350,450
Second group: totally 13175.Amplitude (V): 1-4 (step:0.1);Frequency (Hz): 2,10,30,50,70,90,110,
130-185(step:5),200-250(step:10);Pulsewidth (μ s): 30-120 (step:10), 150-450 (step:50).
Third group: totally 756.Amplitude (V): 4.5-10 (step:0.5);Frequency (Hz): 2,30,70,110,130,150,
170,200,250;Pulsewidth (μ s): 30,70,110,150,200,350,450.
The impedance of sample above is 1000 Ω.Impedance and battery power voltage are not being considered with this 14561 samples
Premise goes down to predict more than 530,000 stimulation parameters (amplitude, frequency, pulsewidth) combinations.(only considering voltage mode)
The data of three dimensions, i.e. amplitude, frequency and pulsewidth are acquired using Auto-Test System.Due to increasing a dimension
It can affect to data acquisition and precision of prediction, therefore there is no impedance information is added, can be examined below as correction factor
Consider.
Be explained in detail below why increase a dimension can to data acquire and precision of prediction affect.From number
According to the angle of acquisition, it is now desired to the sample size of acquisition is 14561, if increasing impedance information, such as is added 300,
This five impedance values of 500,1000,2000,3000 (unit Ω), it means that 14561 × 5=72805 will be needed to acquire
Sample, every one sample of acquisition of Auto-Test System need at least time of 30s, only need 5 days data acquisition times originally, add
25 days are at least needed after entering impedance information.It can be seen that increasing a dimension brings certain difficulty to data collection task.From mould
The angle of type precision of prediction, itself for support vector machines, the feature of input sample are more, that is, dimension is higher, SVM
Precision of prediction can be lower.
In the step S14, sample data pretreatment refers to the input feature vector amplitude to sample data, frequency and pulsewidth
Numerical value is handled.Suitable for this problem sample preprocessing mode there are mainly two types of: normalize and take sequence number.Normalization is
More common data prediction mode, and taking sequence number is that the author proposes according to the actual conditions to be studied a question.Referring to
Fig. 6-Figure 11 can be seen that by simulation figure and contrast table and take the processing scheme of sequence number that can make model prediction accuracy significantly
Improve.Normalization scheme is divided into two kinds, one is to sample all input feature vectors and output power consumption be all normalized;
Another is only all input feature vectors of sample to be normalized, and export power consumption then without processing, uses it
Initial data.Image simulation illustrates both different normalization modes, there is biggish difference on model prediction accuracy, the former
Precision of prediction is very low, and the latter's precision of prediction is more much higher than the former, and is just slightly below the scheme for taking sequence number.Separately below to this
Three kinds of sample data pretreating schemes are analyzed, and draw the simulation figure of power consumption prediction, carry out prediction error comparison.
The normalized is to handle according to following normalization mode data:
X1=2 ((X-Xmin)/(Xmax-Xmin))-1 (9)
Wherein, X is former data;X1The value for being X after normalized;XmaxAnd XminData where respectively former data X
Maximum value and minimum value in group.
The method for taking sequence number are as follows: amplitude sequence number=amplitude × 20;Pulsewidth sequence number=(pulsewidth -30)/10;Frequently
Rate sequence number is numbered according to from natural number 0-61 with 1 for increment.As shown in table 1 below.
Table 1 takes sequence number to pre-process input sample characteristic value
Referring to table 2, for the mode model prediction accuracy pair for being respectively adopted not to sample preprocessing, normalizing and taking sequence number
Than:
Table 2
Standard deviation
Wherein, xiFor predicted value, x is actual value, and n is test set sample number.
By to above 3 group of 6 width Matlab analogous diagram and using pair of different sample preprocessing scheme model prediction accuracies
Than table 2, it can be deduced that such as draw a conclusion:
First, in algorithm of support vector machine, if do not pre-processed to trained and forecast sample, model prediction accuracy
Can be excessively poor, or even can not show a candle to least square regression model.Such as in Fig. 6 and Fig. 7 and model prediction accuracy contrast table 2
Two column.
Second, by Fig. 8-Figure 11, it is most suitably used for taking the mode of sequence number to sample input feature vector (amplitude, frequency, pulsewidth)
In the sample preprocessing scheme that this paper is studied a question, by taking sequence number that can make model prediction accuracy highest.To its reason into
Row brief analysis: 1. amplitude is equal proportion amplification, becomes original 10 times;2. processing is numbered in frequency, larger data is reduced
While value, do not become smaller compared with small data, the gap between data reduced while guaranteeing relationship between data,
And the rule of the internal feature between data and connection are enhanced, while data area being made to become smaller and more effectively, data distribution obtains
More rationally;3. a linear transformation pw=30+10 × n (pw is practical pwm value, and n is corresponding sequence number) has been done to pulsewidth,
Since practical turnable pulse width value is between 30-450us, data value can be made to effectively reduce, reduce when do not break yet data it
Between internal feature relationship.And not 10 times are directly reduced, but has an intercept, thus ensure that the integrality of data,
So that transformed data are since 0, and numerical value is also smaller.To sum up, to the input feature vector of sample by taking sequence number in this way
Mapping relations can completely save the relationship between different characteristic value inside, without break in characteristic value association.
In addition, taking sequence number according to simulation figure and only all input feature vectors of sample being normalized, and export power consumption not
Carry out processing using both sample preprocessing schemes of its initial data model prediction accuracy it is suitable, due to length etc. this
In do not provide simulation figure comparison.
Third, normalization scheme in Figure 10 and Figure 11 using to sample all input feature vectors and output power consumption all
The mode being normalized.It can be seen that this sample preprocessing scheme is compared to taking sequence number precision of prediction slightly poor.It is right
Its reason carries out brief analysis.This processing mode and only all input feature vectors of sample are normalized, and are exported
Power consumption then has a very big difference using the scheme of its initial data without handling, and is exactly this mode due to training sample
This output power consumption is normalized, then it is also normalization that the power consumption predicted value obtained after the trained study of SVM, which is equivalent to,
Afterwards, it at this moment just needs to carry out renormalization to it.I.e. if first by entire data set (including training data and prediction data) into
Row normalization, then trains SVM model, then go to give a forecast with model, then obtained prediction result is carried out renormalization, obtains
The power consumption data of prediction.It is had a problem that in this process itself, due to being not aware that predicted value pair when being exactly renormalization
The maximum and minimum value for the power consumption answered, but still actual measurement training data concentrate power consumption maximin, unless this two
The maximin of person is identical, otherwise will necessarily there is deviation.
Therefore, for the training of SVM algorithm, to the prediction effect got well, most important is exactly data, i.e., for original
The feature extraction and processing of beginning data.Model prediction accuracy may be generated and be expected not using different sample preprocessing schemes
The larger impact arrived.Simulation result explanation, the data prediction of sequence number is taken to sample input feature vector (amplitude, frequency, pulsewidth)
Mode can make model prediction accuracy highest, and the standard deviation of forecast sample is the smallest.And it tries not to export sample special
Sign is that power consumption number does any processing (such as: normalization and renormalization), this can make precision of prediction have a greatly reduced quality.
In the step S17, on the basis of selected kernel function, its parameter is in optimized selection.Kernel function
Selection and parameter are preferably to determine the key of SVM regression forecasting precision, this generally requires certain priori knowledge, there is presently no
General conclusion here carries out penalty coefficient C and nuclear parameter γ by the way of grid search preferred.
The objective function of optimization is mean square error mse, and participating in preferred parameter is penalty factor and nuclear parameter γ, preferably
Range are as follows: log2The combination of C=-15:15, log2 γ=- 15:15, C and γ share 961 kinds.Parameter optimization scheme is to utilize net
The mode of lattice search selects to make the smallest one group of parameter of mean square error mse as optimal C and γ.At best parameter group,
The empiric risk and fiducial range of SVM be both not in close to optimal combination (in structural risk minimization schematic diagram 2 at h*)
Overfitting phenomenon will not occur owing study phenomenon, and SVM has good Generalization Ability at this time.Meanwhile it being generated at random when training
Training sample and test sample, and 3 folding cross validations and parameter are preferably combined, to prevent over-fitting, improve the general of model
Change ability.
Figure 12 is the preferred schematic diagram of certain primary parameter (stimulation amplitude 0-0.9V, impedance 1000 Ω, supply voltage 2.8V;
Sample data pretreatment mode is using the scheme for taking sequence number).Abscissa is log2C in figure, and ordinate is log2 γ, different face
The line of color indicates mean square error mse isopleth, indicates that mse is bigger closer to red (warm colour), indicates closer to blue (cool colour)
Mse smaller (having the specific value for marking mse in figure).The position irised out in Figure 12 is log at this time at mse minimum2C=14,
log2γ=- 10, i.e. best C=16384, γ=9.76563 × 10 best-4。
Be brain pacemaker electric quantity monitoring system 10 provided by the invention referring to Figure 13 comprising: a control module 100 with
And computing module 101, data acquisition module 102, the communication module 103 being connect with the control module 100, display module 104, number
According to input module 105 and memory module 106.
The control module 100 controls the work of entire brain pacemaker electric quantity monitoring system 10.The data acquisition module
102 are used to obtain the stimulation parameter of the brain pacemaker, such as: amplitude amp, pulsewidth pw, impedance im and frequency cf.It is described
Data acquisition module 102 can be communicated by communication module 103 and the vitro program controlled instrument of the brain pacemaker, to obtain
The stimulation parameter of brain pacemaker.The data input module 105 is for inputting battery total capacity, battery from buying process finishing
Time and brain pacemaker from factory to information such as the times of booting.The computing module 101 is used to pass through above-mentioned formula root
Power consumption, remaining capacity and the service life of the brain pacemaker are calculated according to stimulation parameter.The display module 104 is by the meter
Calculate that module 101 is calculated as the result is shown to user.Referring to Figure 14, the display module 104 passes through a graphic user interface
Monitoring data is shown to user, wherein display data include: battery total capacity, battery from buy process finishing time,
Brain pacemaker is from factory to the time of booting, power consumption, remaining capacity, error and stimulation parameter amplitude amp, pulsewidth pw, impedance
Im and frequency cf.The memory module 106 information for storing data.
Further, the brain pacemaker electric quantity monitoring system 10 can also not have data acquisition module 102 and communication module
103, completely by data input module 105 come hand input-data information.
Further, the brain pacemaker electric quantity monitoring system 10 can also be by communication module 103 by the computing module
101 results being calculated are sent to the clients such as mobile phone.
Further, the brain pacemaker electric quantity monitoring system 10 can also include a comparison module 107 and a cue module
108.The comparison module 107 is used to for the result that the computing module 101 is calculated being compared with a secure threshold, from
And it is prompted in danger by the cue module 108.
The present invention includes: to obtain brain pace-making using the method that the electric quantity monitoring system of above-mentioned brain pacemaker carries out electric quantity monitoring
The stimulation parameter of device;The battery relevant parameter of the brain pacemaker is calculated by SVM prediction model;And it will calculate
To battery parameter be shown to user.The battery relevant parameter includes power consumption, electric quantity consumption, remaining capacity and battery life
One of or it is a variety of.
Multiple embodiments of the invention are given above, it is to be understood that without departing from present disclosure essence
In the case where mind and range, it can make a variety of changes, replace and change, these embodiments are also in guarantor of the invention
It protects in range.
Claims (5)
1. a kind of electric quantity monitoring method of brain pacemaker, this method comprises:
Obtain the stimulation parameter of brain pacemaker;
The battery relevant parameter of the brain pacemaker is calculated by SVM prediction model;And
The battery parameter being calculated is shown to user;
Wherein, the SVM prediction model is established by the following method:
Step S10 determines selected kernel function, enters step S11;
Step S11 initializes the parameter of kernel function, enters step S12;
Step S12 acquires the data of three dimensions using Auto-Test System, and amplitude, frequency and pulsewidth are to obtain experiment sample
Notebook data enters step S13;
Sample data is used to learn or predict to compare, if it is study, enters step S14 by step S13, if it is prediction pair
Than then entering step S18;
Step S14, pre-processes sample data, and described is to take sequence number to the pretreated method of sample data progress, institute
State the method for taking sequence number are as follows: amplitude sequence number=amplitude × 20;Pulsewidth sequence number=(pulsewidth -30)/10;Frequency sequence number is pressed
It is numbered according to natural number with 1 for increment, enters step S15;
Step S15 judges whether selected parameter attribute is representative, if so, S16 is entered step, if it is not, then returning to step
Rapid S14;
Step S16, algorithm training establish model, enter step S17;
Step S17 enters step S18 using trellis search method Optimal Parameters;
Step S18, prediction data and experimental data are compared, and enter step S19;
Step S19, error in judgement whether within the allowable range, if so, S20 is entered step, if it is not, then return step S16;
And
Step S20, as prediction model.
2. the electric quantity monitoring method of brain pacemaker according to claim 1, which is characterized in that in the step S10, institute
The kernel function of choosing are as follows: Radial basis kernel function K (xi,xj)=exp (- γ | | xi-xj||2),γ>0。
3. a kind of brain pacemaker electric quantity monitoring system comprising:
One control module, the control module control the work of entire brain pacemaker electric quantity monitoring system;
One computing module, the computing module are used to calculate the brain by any one method described in the claims 1 to 2
The battery relevant parameter of pacemaker;
One display module, the display module be used to for the computing module being calculated as the result is shown to user;
One data input module, the data input module are used to input the data information of brain pacemaker;And
One memory module, memory module information for storing data.
4. brain pacemaker electric quantity monitoring system according to claim 3, which is characterized in that further comprise a data acquisition
Module and a communication module;The data acquisition module is used for the vitro program controlled instrument by the communication module and the brain pacemaker
It is communicated, to obtain the stimulation parameter of brain pacemaker.
5. brain pacemaker electric quantity monitoring system according to claim 3, which is characterized in that further comprise a comparison module
With a cue module;The comparison module is for comparing the result that the computing module is calculated with a secure threshold
Compared with to be prompted in danger by the cue module;And the cue module will be described by a graphic user interface
Computing module be calculated as the result is shown to user.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610629704.4A CN106249159B (en) | 2016-08-03 | 2016-08-03 | A kind of electric quantity monitoring method and electric quantity monitoring system of brain pacemaker |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610629704.4A CN106249159B (en) | 2016-08-03 | 2016-08-03 | A kind of electric quantity monitoring method and electric quantity monitoring system of brain pacemaker |
Publications (2)
Publication Number | Publication Date |
---|---|
CN106249159A CN106249159A (en) | 2016-12-21 |
CN106249159B true CN106249159B (en) | 2019-01-29 |
Family
ID=57606448
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201610629704.4A Active CN106249159B (en) | 2016-08-03 | 2016-08-03 | A kind of electric quantity monitoring method and electric quantity monitoring system of brain pacemaker |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106249159B (en) |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102680907A (en) * | 2012-05-31 | 2012-09-19 | 浙江大学 | Battery charging stress optical coefficient (SOC) detection method in photovoltaic system |
CN203798989U (en) * | 2014-04-22 | 2014-08-27 | 桂林电子科技大学 | SOC detection device for power battery pack of electromobile |
CN104182630A (en) * | 2014-08-20 | 2014-12-03 | 国家电网公司 | Residual battery capacity detection method based on simplified least square support vector machine |
CN105116343A (en) * | 2015-08-24 | 2015-12-02 | 桂林电子科技大学 | LS-SVM power cell SOC estimation method and system |
CN105470593A (en) * | 2015-12-25 | 2016-04-06 | 惠州市德赛电池有限公司 | System for reading power management data via single-chip microcomputer IIC communication |
CN105678077A (en) * | 2016-01-07 | 2016-06-15 | 北京北交新能科技有限公司 | Online prediction method of power performance of lithium ion battery for hybrid power vehicle |
-
2016
- 2016-08-03 CN CN201610629704.4A patent/CN106249159B/en active Active
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102680907A (en) * | 2012-05-31 | 2012-09-19 | 浙江大学 | Battery charging stress optical coefficient (SOC) detection method in photovoltaic system |
CN203798989U (en) * | 2014-04-22 | 2014-08-27 | 桂林电子科技大学 | SOC detection device for power battery pack of electromobile |
CN104182630A (en) * | 2014-08-20 | 2014-12-03 | 国家电网公司 | Residual battery capacity detection method based on simplified least square support vector machine |
CN105116343A (en) * | 2015-08-24 | 2015-12-02 | 桂林电子科技大学 | LS-SVM power cell SOC estimation method and system |
CN105470593A (en) * | 2015-12-25 | 2016-04-06 | 惠州市德赛电池有限公司 | System for reading power management data via single-chip microcomputer IIC communication |
CN105678077A (en) * | 2016-01-07 | 2016-06-15 | 北京北交新能科技有限公司 | Online prediction method of power performance of lithium ion battery for hybrid power vehicle |
Non-Patent Citations (2)
Title |
---|
磷酸铁锂电池荷电状态估算方法研究;李智鹏;《中国优秀硕士学位论文全文数据库 工程科技Ⅱ辑》;20131015(第10期);正文第43,47-48页,图5.1 |
脑起搏器与神经调控技术;李路明;《生命科学仪器》;20150625;第13卷(第3期);第7-11页 |
Also Published As
Publication number | Publication date |
---|---|
CN106249159A (en) | 2016-12-21 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Liu et al. | A deep learning method for prediction of three‐dimensional dose distribution of helical tomotherapy | |
de Florian et al. | Global analysis of helicity parton densities and their uncertainties | |
CN106093795B (en) | A kind of electric quantity monitoring method and electric quantity monitoring system of brain pacemaker | |
Sujit et al. | Automated image quality evaluation of structural brain MRI using an ensemble of deep learning networks | |
DE102016203546B4 (en) | ANALYZER FOR BEHAVIORAL ANALYSIS AND PARAMETRIZATION OF NEURONAL STIMULATION | |
US20180043182A1 (en) | Systems and methods for automated radiation treatment planning with decision support | |
Wang et al. | Realizing low-energy classification systems by implementing matrix multiplication directly within an ADC | |
Xing et al. | A feasibility study on deep learning‐based radiotherapy dose calculation | |
Ma et al. | Dose distribution prediction in isodose feature‐preserving voxelization domain using deep convolutional neural network | |
Liu et al. | A deep learning model to predict dose–volume histograms of organs at risk in radiotherapy treatment plans | |
Zhan et al. | A glioma segmentation method using cotraining and superpixel-based spatial and clinical constraints | |
Zhu et al. | Search for IMRT inverse plans with piecewise constant fluence maps using compressed sensing techniques | |
Rudra et al. | Forecasting of breast cancer and diabetes using ensemble learning | |
Jiang et al. | Convolutional neural network‐based dosimetry evaluation of esophageal radiation treatment planning | |
Lizar et al. | Patient-specific IMRT QA verification using machine learning and gamma radiomics | |
Wang et al. | Patient feature based dosimetric Pareto front prediction in esophageal cancer radiotherapy | |
CN106249159B (en) | A kind of electric quantity monitoring method and electric quantity monitoring system of brain pacemaker | |
Kalantzis et al. | Toward IMRT 2D dose modeling using artificial neural networks: a feasibility study | |
Xue et al. | Online capacity estimation of lithium-ion batteries based on deep convolutional time memory network and partial charging profiles | |
Gao et al. | A hybrid cnn-bilstm approach for remaining useful life prediction of evs lithium-ion battery | |
Fan et al. | Combining model-and deep-learning-based methods for the accurate and robust segmentation of the intra-cochlear anatomy in clinical head CT images | |
Kump et al. | An automated treatment plan alert system to safeguard cancer treatments in radiation therapy | |
CN113033064B (en) | Power equipment impact load simulation method based on support vector machine | |
Hao et al. | Improved coyote optimization algorithm for parameter estimation of lithium-ion batteries | |
CN106250633B (en) | A kind of electric quantity monitoring method and electric quantity monitoring system of brain pacemaker |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |