CN101894185A - Method for predicting life of small sample data object based on dynamic bipolar modified probabilistic neural network (MPNN) - Google Patents
Method for predicting life of small sample data object based on dynamic bipolar modified probabilistic neural network (MPNN) Download PDFInfo
- Publication number
- CN101894185A CN101894185A CN2010102209511A CN201010220951A CN101894185A CN 101894185 A CN101894185 A CN 101894185A CN 2010102209511 A CN2010102209511 A CN 2010102209511A CN 201010220951 A CN201010220951 A CN 201010220951A CN 101894185 A CN101894185 A CN 101894185A
- Authority
- CN
- China
- Prior art keywords
- life
- data
- span
- mpnn
- prediction
- 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
Images
Abstract
The invention discloses a method for predicting the life of a small sample data object based on a dynamic bipolar modified probabilistic neural network (MPNN) and belongs to the technical field of the life prediction of small samples. The method comprises the following steps of: obtaining a function relation between life influencing factors and life characteristic parameters of a prediction object by collecting all available data of a life prediction object and preprocessing life prediction related data and analyzing data dependency for the life prediction object; then mapping data and obtaining equivalent data values of the prediction object; performing a primary MPNN training and prediction and a secondary MPNN training and prediction by using the improved MPNN; and finally, determining life values of the prediction object according to a life termination criterion of the life characteristic parameters of the prediction object. The method can realize the life prediction of the small sample data object by fully using the life influencing factor data of the prediction object under an adverse condition of a little practical available data of the prediction object and has a strong universality; and moreover, a life prediction model can dynamically regulate a neural network and can constantly improve the precision of the life prediction.
Description
Technical field
The invention belongs to small sample forecasting technique in life span field, specifically, be meant a kind of life-span prediction method that is applied to the small sample data object.
Background technology
Scope that forecasting technique in life span relates to and field are extremely extensive, from raw-material fatigue lifetime to the complicated shaped article life-span, all need forecasting technique in life span from civil area to the national defence field.At present, carry out life prediction work and mainly contain following three kinds of methods:
A, based on the life prediction of physical model: this method is based on the analysis to research object, thereby sets up the physical model of reflection object evolution process, by related data model parameter adjusted, and obtains the life prediction model that needs at last.This method is mainly used in material science;
B, based on the life prediction of statistical model hypothesis: these class methods suppose that at first the research object life-span obeys certain statistical distribution, utilize a large amount of existing lifetime datas to determine the parameter of this model afterwards, thereby set up the life prediction model of research object;
C, life prediction based on life-span analysis of Influential Factors training: the method mainly is each the life-span influence factor by research and definite impact prediction object, and set up influence factor with the relational network between the life-span, thereby set up the life prediction model of research object by a large amount of testing data of life-span.
For above-mentioned three kinds of methods, need further investigate the internal mechanism of forecasting object based on the life prediction of physical model, for complication system, the workload of setting up corresponding physical model will be very huge; The life prediction that reaches based on life-span analysis of Influential Factors training based on the statistical model hypothesis then needs a large amount of object lifetime data to set up the life prediction model.Consider in practical engineering application,, tend to run into various objective condition restrictions, can not have a large amount of lifetime datas that is used for life prediction especially at aerospace field.Thereby, to study a kind ofly at few lifetime data, the life-span prediction method with small sample data characteristics has great importance concerning the object that meets these characteristics.
(Artificial Neural Network is a simulation cerebral nervous system 26S Proteasome Structure and Function ANN) to artificial neural network, is the artificial network that neuron extensively connects to form by a large amount of simple process unit.It can be from given data induction rule automatically, obtain the inherent law of these data, have very strong non-linear mapping capability.Artificial neural network has following outstanding advantage: 1. Gao Du concurrency; 2. the non-linear overall situation effect of height; 3. good fault-tolerance and function of associate memory; 4. ten fens strong self-adaptations, self-learning function.
Probabilistic neural network (being called for short the PNN network) is a kind of of artificial neural network, the topological structure of probabilistic neural network model as shown in Figure 1, probabilistic neural network is a three-decker, is respectively input layer, radial basis function layer (also being hidden layer) and competitive learning layer (also being output layer).Wherein, on behalf of input layer, R the input vector of R component is arranged, and N is the number of input training sample vector,
yBe input vector, b
[1]Be the threshold vector of hidden layer, K is output layer neuron number (i.e. Fen Lei a classification number),
w [1]With
w [2]Be respectively the neuronic weight matrix of hidden layer and output layer, n
[1]And n
[2]Be respectively the neuronic weighted sum vector of hidden layer and output layer, a
[1]And a
[2]Be respectively the neuronic output vector of hidden layer and output layer, its expression formula is respectively:
a
[1]＝radbas(||
w [1]-
y||·*b
[1])
a
[2]＝compet(w
[2]·a
[1])
Radial basis function is in the PNN network: Radbas (u)=exp (u
2)
The training of this PNN network and assorting process can be expressed as: the hidden layer weight matrix
w [1]Be changed to input training sample matrix Y
(N * R), the output layer weight matrix
w [2]Be changed to the objective matrix T of training sample
(K * N)Each row of objective matrix have only one 1, and remaining is 0.When importing the sample vector X that will classify, hidden layer calculates the distance between input vector and the training sample, and the probability density of each classification of output ownership; The competitive learning excitation function of output layer is selected the maximum probability density of corresponding classification, and corresponding generation 1, and all the other classifications are 0.
The PNN network has that network training speed is fast, precision of prediction is high, need not advantages such as repetition training for new data, is widely used in the practical engineering project, especially for classification problem.Yet for forecasting problem, the PNN network seems at a loss what to do, thereby can't directly utilize the PNN network to carry out life prediction to having small sample data characteristic object more.
Summary of the invention
In the engineering reality of reality, can not obtain traditional life prediction desired data under a lot of situations.Consider the characteristics that need plenty of time sequence and lifetime data in traditional forecasting technique in life span, for the forecasting object that is difficult to obtain data sample, research is making full use of under the existing information prerequisite, predicts that as far as possible accurately the residual life value is a problem that has challenge.The present invention is directed to the deficiency that exists in traditional forecasting technique in life span, take all factors into consideration the relative merits of PNN network, on to the improved basis of PNN network, make up dynamic bipolar MPNN network model, proposition utilizes dynamic bipolar MPNN method, realizes the life prediction to the forecasting object with small sample data characteristics.The dynamic whole M PNN network of adjusting guarantees that in whole life prediction process precision of prediction is along with the prolongation of time and the increase of data volume improve constantly in the process of life prediction.
Life-span prediction method provided by the invention is specifically realized as follows:
All data availables of step 1, collection life prediction object;
By to forecasting object and like product analysis, collect utilizable all life prediction related datas.
The life prediction data that step 1 obtains are analyzed and screened, need extract life-span characterization parameter and the life-span influence factor data that can be used for dynamic bipolar MPNN life prediction according to data of the present invention.Simultaneously, the data that filter out are carried out pre-service work such as singular value rejecting, data noise reduction.
The life-span influence factor data that screen in the step 2 are carried out correlation analysis, thereby obtain these life-span influence factors with funtcional relationship between forecasting object life-span characterization parameter.
Described correlation analysis is meant by approximation of function or utilizes SPSS that realization is analyzed the different parameters correlation of data, thereby obtains the mapping relations between supplemental characteristic.
Utilize that step 3 obtains under each life-span influential factors, the like product that obtains is with the correlationship between forecasting object life-span characterization parameter, based on the like product data of reference, shine upon and obtain the equivalent data value of forecasting object life-span characterization parameter.
According to the life prediction demand, classification PNN network model is improved, and obtain prediction MPNN network model.
The foundation of step 6, a MPNN network, training and prediction;
Determine that a MPNN imports node and output node number.Utilize forecasting object life-span characterization parameter and do the processing of difference ratio by the equivalent data that step 4 obtains, training sample and the test sample book of a MPNN of structure; In order to reject the singular value in the training sample, accelerate the speed of convergence of network, to carry out normalized by input vector, the object vector of above-mentioned structure, importing a MPNN network then trains it, thereby determine a MPNN network parameter, and utilize a MPNN neural network that trains, the life-span characterization parameter is predicted.
Foundation, training and the prediction of step 7, secondary MPNN network;
Determine that secondary MPNN determines input number of nodes and output node number respectively.Utilization forecasting object after pre-service characterizes predicting the outcome of life parameter data and a MPNN, training and the test sample book of structure secondary MPNN, and then carry out prediction work, determine the life value of forecasting object at last according to the end-of-life criterion of forecasting object life-span characterization parameter.
Step 8, dynamic time window are adjusted;
V is according to practical object characteristics and request for utilization, is set as one month, two weeks, a week etc. the corresponding dynamic time interval.According to the time interval that is provided with, repeat above-mentioned steps two afterwards, bipolar MPNN network is trained again and predicted to step 7.
Advantage of the present invention is:
(1) life-span prediction method provided by the invention can make full use of forecasting object life-span influence factor data under actual prediction object data available adverse condition seldom, realizes the life prediction of small sample data object;
(2) because the present invention makes up and forms, so in the life prediction process, need not to set up the life-span influence factor and forecasting object life-span characterization parameter concerns with the analytical function between the life-span on the PNN basis;
(3) life-span prediction method that proposes of the present invention is the solution solution at the class problem with small sample data characteristics, has stronger versatility;
(4) dynamic bipolar MPNN network is based on the PNN network, through improve and structure after and the life prediction model of formation, thereby it is keeping the quick training of PNN network and the advantage of prediction; The dynamic bipolar MPNN method that the present invention proposes has remedied the problem and shortage that traditional forecasting technique in life span exists, thereby solves small sample life prediction problem, and then realizes the once leap of forecasting technique in life span from theoretical research to the engineering practical application.
(5) for particular problem, life prediction model of the present invention can dynamically be adjusted neural network, can constantly promote life prediction precision.
Description of drawings
Fig. 1 is a probabilistic neural network topological structure synoptic diagram in the prior art;
Fig. 2 is through improved probabilistic neural network topological structure synoptic diagram among the present invention;
Fig. 3 is the overview flow chart of life-span prediction method provided by the invention;
Fig. 4 is raw data figure;
Fig. 5 is ground data figure behind exceptional value processing and noise reduction;
Fig. 6 is the data map figure of correlation analysis;
Fig. 7 is a MPNN prediction measuring accuracy curve map;
Fig. 8 is a MPNN life prediction performance degradation curve map;
Fig. 9 is used for life of storage battery prediction effect figure for dynamic bipolar MPNN.
Embodiment
The present invention is described in detail below in conjunction with drawings and Examples.
The present invention is a kind of life-span prediction method of the small sample data object based on dynamic bipolar MPNN network, described life-span prediction method is a kind of nonparametric technique, this method does not need to set up the life-span influence factor and historical data concerns with the parametric function between the object lifetime, only need to obtain the MPNN network after the PNN network improves to having now, through after the analysis and pre-service work to forecasting object, the data that utilization is handled constitute the training set and the forecast set of a MPNN network, through after the training study, utilize a MPNN network prediction to replenish historical data sample, after finishing these work, promptly can utilize secondary MPNN network to carry out life prediction work, the end-of-life time is determined in judgement according to end-of-life at last, and then obtain needed life prediction value, Figure 3 shows that the overview flow chart of life-span prediction method of the present invention, concrete implementation step is as follows:
By to forecasting object and like product analysis thereof, collect utilizable all life prediction related datas.As: for accumulator, its data available is: temperature, charging current, discharge current, discharge electricity amount, load current data, charge capacity, dump energy, voltage, discharge and recharge than etc. supplemental characteristic; For solar battery array, its data available is: supplemental characteristics such as temperature, square formation current data, load current data, power, busbar voltage, stepup transformer output voltage.
Described like product is meant on physical arrangement, logical organization and functional structure, with the similar or identical product of forecasting object.
The life prediction related data that obtains in the step 1 is analyzed and screened, extract the life-span characterization parameter and the life-span influence factor of forecasting object, and the life-span influence factor is classified.
The life-span characterization parameter (L_index) and the supplemental characteristic of described forecasting object are as follows:
(1) the life-cycle data of like product life-span characterization parameter (L_sim) are promptly brought into use to this similar from like product
The historical time sequence data of life of product termination, its time sequence data can be expressed as: L_sim_x1, L_sim_x2 ..., L_sim_xend;
(2) the existing incomplete lifetime data of forecasting object life-span characterization parameter (L_obj), promptly bring into use up to the present all time series datas from forecasting object, its time sequence data can be expressed as: L_obj_x1, L_obj_x2,, L_obj_xnow.
Described life-span influence factor and life-span influence factor data can be divided into following two classes:
(1) time series life-span influence factor data: have identical time scale life-span influence factor data with forecasting object life-span characterization parameter, these influence factor data can be expressed as P_in1, P_in2, P_in3 ... wherein, P_in1 can be expressed as: P_in1_1, P_in1_2, P_in1_3 ..., P_in1_i ..., P_in1_now, in like manner, P_in2, P_in3 ... also can be expressed as corresponding sequence form.Described life-span influence factor data comprise the life-cycle data and the existing time series data of forecasting object life-span influence factor of like product life-span influence factor, and the form of the composition of like product data and expression way and forecasting object life-span characterization parameter L_index and life-span influence factor are similar;
(2) data are adjusted life-span influence factor data: be different from time series life-span influence factor data, these type of influence factor data are that limited data are right, these data are to being the historical data of like product, reaction be that corresponding life-span influence factor parameter is with the corresponding relation between the life-span.Be used for local forecasting object life-span characterization parameter L_index of adjustment and time sequence P_in_1, P_in2, P_in3 ... data.As: for life of storage battery forecasting object, the depth of discharge parameter is this type of life-span influence factor, and when depth of discharge was 17%, its corresponding life value was 20,000 charge and discharge cycles, constitutes data thus to ((17%-20000)).Such life-span influence factor can be expressed as: P_rel1, P_rel2, P_rel3 ...
After finishing life-span influence factor classification, to life-span characterization parameter L_index, the life-span influence factor data P_in1 of above-mentioned forecasting object, P_in2, P_in3 ... (comprising forecasting object and like product object data) carries out that singular value is rejected, the pre-service of data noise reduction, obtain through the pretreated data L_index_p of data, P_in1_p, P_in2_p, P_in3_p, Wherein, time series L_index_p can be expressed as L_index_p_1, L_index_p_2 ..., L_index_p_now; In like manner, time series P_in1_p, P_in2_p, P_in3_p ... also can be expressed as corresponding sequence form.In the process of carrying out life prediction, the life-span influence factor data of forecasting object and in forecasting process to the operation of these data, must be corresponding one by one with the life-span influence factor data and the operation thereof of similar product.
Consider life-span influence factor P_rel1, P_rel2, the P_rel3 of like product and life prediction object, need the pretreated parameter of like product and forecasting object correspondence to be carried out correlation analysis, thereby obtain the correlationship on the life-span characterization parameter between like product and the forecasting object by approximation of function or SPSS methods such as (statistics product and service solution---Statistical Product and Service Solutions).As: for the historical depth of discharge data of accumulator to (the like product data (and 17%-20000,30%-16000 ... )), by these data to setting up the like product depth of discharge with the funtcional relationship between the life-span.Utilize this funtcional relationship, set up like product the correlationship (as: linear relationship) life-span between of forecasting object life-span with reference according to the actual discharge degree of depth (being assumed to be 20%) of forecasting object and the like product actual discharge degree of depth of reference (supposing that depth of discharge is 30%).
Utilize the correlationship on the life-span characterization parameter between like product that step 3 obtains and the forecasting object,, shine upon and obtain the equivalent data value of forecasting object life-span characterization parameter based on the like product data.As: { correlationship that m value L_index_pm among the L_index_pi} obtains by step 3 is mapped to like product life-span characterization parameter, and { n value L_sim_pn among the L_sim_pi}, then L_sim_pn is the equivalent data value of L_index_pm to suppose the life-span characterization parameter time series of forecasting object.So, can obtain forecasting object life-span characterization parameter sequence { all equivalent data values of L_index_pi}.
As classification tool, probabilistic neural network (PNN network) adopts the kernel function of competitive function (Compet (g)) as the output layer of network.As forecasting tool, then need PNN network output layer kernel function is changed to mathematical expectation function (Expect (g)), be about to Compet (g) and change to Expect (g), just make original classification tool convert forecasting tool to, and the characteristics of original statistics stability and training rapidity have been kept, obtain the MPNN network thus, as shown in Figure 1 and Figure 2.By bayesian theory as can be known, certain class statistical distribution is obeyed in the output of probabilistic neural network.Can accept above-mentioned modification like this, promptly replace competitive function and make kernel function, thereby make the statistical tool of probabilistic neural network have forecast function with the mathematical expectation function.So, the training of MPNN network and forecasting process can be expressed as: the hidden layer weight matrix
w [1]Be changed to input training sample matrix Y
(N * R), the output layer weight matrix
w [2]Be changed to the objective matrix T of training sample
(K * N)When importing the sample vector X that will classify, hidden layer calculates the distance between input vector and the training sample, and the probability density of each classification of output ownership; The mathematical expectation function of output layer (Expect (g)) calculates mathematical expectation according to the probability density of upper strata output, and the output of resulting mathematical expectation as the MPNN network.
The foundation of step 6, a MPNN network, training and prediction;
The equivalent data value L_index_p_rel of the forecasting object that obtains to the forecasting object life-span characterization parameter L_index_p value that obtained by step 2 with by step 4 carries out the difference ratio to be handled, and as a MPNN network output.To like product parameter P_sim_in1_p, P_sim_in1_p, the P_sim_in1_p that obtains by step 2, value and forecasting object parameter P_in1_p, P_in1_p, P_in1_p, carrying out the difference ratio handles, as a MPNN network input parameter, and construct the input vector of a MPNN network.So far, construct the input vector and the object vector of a MPNN network; In order to realize rejecting the singular value in the training sample, accelerate the speed of convergence of network, will carry out normalized by the input vector and the object vector of above-mentioned structure; Importing a MPNN network then trains it; At last, utilize a MPNN neural network that trains, the index of aging parameter is predicted.After anti-normalization, obtain the predicted value sequence L_index_pre1_1 of a MPNN, L_index_pre1_2 ..., L_index_pre1_end.Wherein, L_index_pre1_end is the anti-normalized value of MPNN available last data of prediction.
Foundation, training and the prediction of step 7, secondary MPNN network;
According to the experience of forecasting object or checking by experiment, determine the input and the output node number of secondary MPNN network.With { L_index_p} value and the life-span characterization parameter time series forecasting value L_index_pre1_1 that obtains by step 6 through the pretreated forecasting object of step 2, L_index_pre1_2, L_index_pre1_end is the data basis, according to the input of secondary MPNN network and input vector and the object vector of output node number structure secondary MPNN, the line time sequence iteration of going forward side by side prediction; Determine the life-span of forecasting object at last according to the end-of-life criterion of index of aging parameter.
Step 8, dynamic time window are adjusted;
In actual engineering, As time goes on, the data volume of forecasting object on above-mentioned parameter increases gradually, and new data can effectively promote life prediction precision.According to actual prediction features of the object and request for utilization, corresponding dynamic time window value is set, as: one month, two weeks, a week etc.According to the time window value that is provided with, repeat above-mentioned steps two afterwards, bipolar MPNN network is trained again and predicted, redefine the life value of forecasting object to step 7.
The present invention analyzes and organization network input and output data by rational on the basis of modified probabilistic neural network (MPNN network), modified probabilistic neural network of structure respectively stage by stage (MPNN) and secondary modified probabilistic neural network (secondary MPNN), utilize these two modified neural network predictions that make up, and obtain the life value of forecasting object.In addition, the present invention can upgrade the life prediction data under given time window prerequisite, and neural network training again, and then obtains the life value of the forecasting object under the new data.The user can understand the remaining life of forecasting object by life prediction, thereby can control the forecasting object life-span and make a strategic decision by the configuration environment for use foundation is provided for logistics management, make under the prerequisite that the topmost task that guarantees forecasting object is implemented, utilize and use forecasting object to greatest extent, reliably, thereby give full play to the usefulness of forecasting object.
Embodiment
Present embodiment is a forecasting object with China's moonlet accumulator.Because the life-span related data of moonlet power-supply system is few, meets the small sample data life prediction problem that will solve required for the present invention.By elaborating of present embodiment, further specify implementation process of the present invention and engineering application process.
For the moonlet accumulator, the data that can be used for life prediction are: the depth of discharge and 5 pairs of corresponding life values of testing discharge final pressure data (complete data), ground accumulator test discharge current data (constant), ground accumulator test depth of discharge data (constant) and history at rail battery discharging voltage data (fragmentary data), in rail battery discharging current data (fragmentary data), at rail battery discharging depth data (fragmentary data), ground accumulator.
Described " fragmentary data " expression only is the preceding portion of time sequence data of complete lifetime data because still can operate as normal at the rail accumulator.
Utilizing the above-mentioned basis that is used for the data of life prediction, using the life-span prediction method of the present invention's proposition and predicted the step and the method for its application are as follows the life-span of moonlet accumulator:
By the moonlet accumulator is analyzed, it is as follows to collect utilizable all related datas:
The forecasting object data---at rail moonlet accumulator:
(1) at rail battery discharging voltage data (fragmentary data);
(2) at rail charge in batteries voltage data (fragmentary data);
(3) in rail battery discharging current data (fragmentary data);
(4) in rail charge in batteries current data (fragmentary data);
(5) at rail battery discharging electric quantity data (fragmentary data);
(6) at the rail accumulator cell charging and discharging than (fragmentary data);
Like product data---the similar model accumulator in ground:
(1) ground accumulator test discharge final pressure data (complete data);
(2) ground accumulator test discharge current data (constant);
(3) ground accumulator test charge current data (constant);
(4) ground accumulator test depth of discharge data (constant);
(5) Li Shi depth of discharge and corresponding life value are 5 pairs;
All related datas that obtain in the step 1 are reasonably analyzed, in conjunction with the above-mentioned description of this invention, extract can be used for dynamic bipolar MPNN network life characterization parameter and the influence factor data as follows:
(1) life-span characterization parameter:
In rail battery discharging final pressure data ({ L_index})---can extract by ' at rail moonlet battery discharging voltage ' and obtain;
Ground accumulator test discharge final pressure data (L_sim});
(2) life-span influence factor:
The accumulator cell charging and discharging cycle index (be included in rail P_in1} and ground P_sim_in1})---can obtain (being about 100.7min weekly working time phase) from experiment and in orbit time at the rail moonlet;
In rail battery discharging current data ({ P_in2});
Ground accumulator test discharge current data (P_sim_in2});
At rail battery discharging depth data (P_rel1)---can extract acquisition by ' at rail battery discharging electric quantity data ';
Ground accumulator test depth of discharge data (P_sim_rel1);
5 pairs of historical depth of discharge and corresponding life values;
To testing discharge final pressure data at rail battery discharging voltage data, ground accumulator and carrying out singular value in rail battery discharging current data and reject, data noise reduction pre-service work.Fig. 4, Fig. 5 handle and ground, noise reduction front and back accumulator test discharge final pressure data plot for picking through singular value.Contrast as can be seen from figure, the data behind the value of picking and noise reduction are more regular, undulatory property is littler.In addition, need be from rail battery discharging voltage, obtaining in rail battery discharging final pressure data, with performance and the life status that is characterized in the rail accumulator.At this moment, the discharge final pressure that obtains is forecasting object life-span characterization parameter L_index.
After the step 2 pre-service, the life-span influence factor that remains is respectively: battery discharging electric current, the battery discharging degree of depth and accumulator cycle charge-discharge number of times.Wherein, discharge current and charge and discharge cycles number of times are as the network input parameter of dynamic MPNN, in order to reflect its influence to the life-span; The influence of depth of discharge can get by the correlation analysis of depth of discharge.
At first, the depth of discharge and the corresponding life value of 5 pairs of history by the similar model accumulator in ground (like product) that obtained by step 1 are set up the like product depth of discharge with the funtcional relationship L_sim=f between the life-span (P_sim_rel1).L_sim=f in the present embodiment (P_sim_rel1) is a piecewise function, and a preceding part is a nonlinear function, and a back part is a linear function.According to this funtcional relationship, ground experiment depth of discharge data (30%DOD) and in the substitution of the rail battery discharging degree of depth (17%DOD) data, can obtain corresponding life value, and then obtain the ground experiment accumulator and coexist that correlationship between the rail life of storage battery---both are the linear ratio relation in the present embodiment, suppose for convenience of description: in the rail life-span: life-span=2, ground herein: 1, and only consider one of DOD " data are adjusted the life-span influence factor " in the present embodiment.
Described DOD is a depth of discharge;
Utilize between the ground accumulator that step 3 obtains 2: 1 correlationship on the life-span characterization parameter,, shine upon and obtain equivalent data value in rail battery discharging final pressure based on ground battery discharging final pressure data.As shown in Figure 6, owing to consider from the angle of DOD, both life-span ratio is 2: 1, that is to say from degree of injury, is equivalent at the degree of injury of rail 17%DOD after discharging and recharging for 2n time with the degree of injury of ground experiment 30%DOD through putting after discharging and recharging for n time, so, obtaining in the equivalent value of the 2n time discharge of rail final pressure value is the n time discharge final pressure value of ground experiment data, thereby finishes the mapping relations of forecasting object with similar product.
The building process of MPNN network is with reference to " summary of the invention " step 5.
Because a MPNN network only has two parameters in the present embodiment: discharge current and charge and discharge cycles number of times.The input of secondary MPNN and output node number are respectively 4 and 1 in the present embodiment.
The foundation of step 6, a MPNN network, training and prediction;
Because in the present embodiment, time series life-span influence factor has only two: battery discharging electric current and charge and discharge cycles number of times.Thereby the input number of nodes of a whole MPNN network is 2, and the output node number is 1.Utilization by step 2 obtain { L_index_p} and done the difference ratio in rail battery discharging final pressure equivalent data and handled by what step 4 obtained, the gained result is as the training output of a MPNN network in rail battery discharging final pressure data through pretreated.Simultaneously, utilize by step 2 obtain { P_in2_p} does the difference ratio with ground battery discharging current data (constant) and handles, and the gained result is as the input of a MPNN network in rail battery discharging current data.In addition, directly the input node of accumulator cell charging and discharging cycle index as a MPNN.So, constitute the input vector of a MPNN by discharge current and charge and discharge cycles number of times.Before carrying out network training to discharge current difference ratio data and discharge final pressure difference ratio The data linear normalization method; Cycle index utilizes the arc tangent method to carry out normalized.Finish training network, and utilize a MPNN network training, predict follow-up in rail battery discharging final pressure value.
Training time: 1.0156s
Precision of prediction: square error MSE=0.0221
Figure 7 shows that MPNN prediction measuring accuracy curve map one time, as can be seen from the figure a MPNN has very high precision in carrying out pattern-recognition and prediction.
MPNN network be output as after the difference normalization corresponding under the corresponding charge and discharge cycles number of times in rail discharge final pressure value.Fig. 8 is a MPNN life prediction performance degradation curve, coincide with existing measured data trend comparison on general trend.
Foundation, training and the prediction of step 7, secondary MPNN network;
The input node of secondary MPNN network is several 4, and the output node number is 1.Utilization by step 2 obtain rail battery discharging final pressure data L_index_p} and by a MPNN life prediction obtain { L_index_pre} constitutes discharge final pressure time series data, and this time series is carried out normalized in rail battery discharging final pressure predicted data; According to time series forecasting thought this time series is carried out iteration training and prediction afterwards.Fig. 9 is a dynamic bipolar MPNN prediction effect synoptic diagram, and wherein the right side thick line partly is depicted as second iteration and predicts the outcome.According to the life-span criterion of HY-1B moonlet accumulator, the bimetry value that obtains present embodiment is about: 5.21, that is: and 5 years 2 months 15 days 14 hours 24 minutes.
Step 8, dynamic time window are adjusted;
The time window of present embodiment is 1week, when the week age that is provided with arrives, utilizes collected new for rail moonlet accumulator data, repeats above-mentioned steps two and can obtain new life prediction value to the step 7 process.
In industry member, especially novel industry or main equipment, in high precision machine tool, aerospace field, it is few to be faced with the forecasting object sample size, and the health condition of forecasting object is to the extremely important problem of its operational management, the present invention is directed to this problem, theoretically, in fact consider to fully utilize the life-span related data object of being studied is carried out life prediction.The present invention by preliminary test figure for newly grinding or improved forecasting object carries out life-span (performance) assessment, the performance reference of design forecasting object is provided for the deviser, for the further improvement of forecasting object provides foundation, reduce forecasting object development risk to a great extent.
Life-span prediction method of the present invention provides reference for optimizing spare part quantity and maintenance schedule, reduce use and maintenance cost, thereby reduce the forecasting object life cycle cost on the whole, obtain the residual life of parts by forecasting technique in life span, can be to some parts recyclings; For some extremely crucial equipment, might cause a large amount of casualties and property loss in case break down, life-span prediction method of the present invention provides important administration base for the decision maker, thereby avoids these catastrophic incidents to take place to greatest extent.
Claims (5)
1. life-span prediction method based on the small sample data object of dynamic bipolar MPNN is characterized in that:
Step 1, all available life prediction data of collecting the life prediction object;
Step 2, the pre-service of life prediction related data;
The life prediction data that step 1 obtains are analyzed and screened, extract the life-span characterization parameter and the life-span influence factor of forecasting object, and the life-span influence factor classified, simultaneously, the data that filter out are carried out singular value is rejected, the pre-service work of data noise reduction;
Step 3, data dependence analysis;
The life-span influence factor data that screen in the step 2 are carried out correlation analysis, thereby obtain these life-span influence factors with funtcional relationship between forecasting object life-span characterization parameter;
Step 4, data map also obtain the equivalent data value of forecasting object;
Utilize that step 3 obtains under each life-span influential factors, the like product that obtains is with the correlationship between forecasting object life-span characterization parameter, based on the like product data of reference, shine upon and obtain the equivalent data value of forecasting object life-span characterization parameter;
Step 5, MPNN network struction;
PNN network output layer kernel function is changed to mathematical expectation function Expect (g), obtain the MPNN network;
The foundation of step 6, a MPNN network, training and prediction;
Utilize forecasting object life-span characterization parameter and do the processing of difference ratio by the equivalent data that step 4 obtains, training sample and the test sample book of a MPNN of structure; To carry out normalized by input vector, the object vector of above-mentioned structure, importing a MPNN network then trains it, thereby determine a MPNN network parameter, and utilize a MPNN neural network that trains, the life-span characterization parameter is predicted;
Foundation, training and the prediction of step 7, secondary MPNN network;
Utilization forecasting object after pre-service characterizes predicting the outcome of life parameter data and a MPNN, training and the test sample book of structure secondary MPNN, and then carry out prediction work, determine the life value of forecasting object at last according to the end-of-life criterion of forecasting object life-span characterization parameter;
Step 8, dynamic time window are adjusted;
According to practical object characteristics and request for utilization, the corresponding dynamic time interval is set, according to the time interval that is provided with, repeat above-mentioned steps two to step 7, bipolar MPNN network is trained again and predicted.
2. the life-span prediction method of the small sample data object based on dynamic bipolar MPNN according to claim 1, it is characterized in that: described correlation analysis is meant by approximation of function or utilizes SPSS, realization is analyzed the different parameters correlation of data, thereby obtains the mapping relations between supplemental characteristic.
3. the life-span prediction method of the small sample data object based on dynamic bipolar MPNN according to claim 1, it is characterized in that: the training and the forecasting process of described MPNN network are expressed as: the hidden layer weight matrix
w [1]Be changed to input training sample matrix Y
(N * R), the output layer weight matrix
w [2]Be changed to the objective matrix T of training sample
(K * N)When importing the sample vector X that will classify, hidden layer calculates the distance between input vector and the training sample, and the probability density of each classification of output ownership; The mathematical expectation function Expect (g) of output layer calculates mathematical expectation according to the probability density of upper strata output, and the output of resulting mathematical expectation as the MPNN network.
4. the life-span prediction method of the small sample data object based on dynamic bipolar MPNN according to claim 1, it is characterized in that: the life-span characterization parameter and the supplemental characteristic of described forecasting object are as follows:
(1) the life-cycle data of like product life-span characterization parameter are promptly brought into use historical time sequence data to this like product life termination from like product, and its time sequence data is expressed as: and L_sim_x1, L_sim_x2 ..., L_sim_xend;
(2) the existing incomplete lifetime data of forecasting object life-span characterization parameter is promptly brought into use up to the present all time series datas from forecasting object, and its time sequence data is expressed as: L_obj_x1, and L_obj_x2 ..., L_obj_xnow.
5. the life-span prediction method of the small sample data object based on dynamic bipolar MPNN according to claim 1, it is characterized in that: described life-span influence factor and life-span influence factor data are divided into following two classes:
(1) time series life-span influence factor data: have identical time scale life-span influence factor data with forecasting object life-span characterization parameter, these influence factor data be expressed as P_in1, P_in2, P_in3 ... wherein, P_in1 is expressed as: P_in1_1, P_in1_2, P_in1_3 ..., P_in1_i ..., P_in1_now, in like manner, P_in2, P_in3 ... also be expressed as corresponding sequence form;
(2) data are adjusted life-span influence factor data: these type of influence factor data are that limited data are right, these data are to being the historical data of like product, reaction be corresponding life-span influence factor parameter with the corresponding relation between the life-span, be used for local forecasting object life-span characterization parameter L_index of adjustment and time sequence P_in1, P_in2, P_in3 ... data.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2010102209511A CN101894185A (en) | 2010-06-29 | 2010-06-29 | Method for predicting life of small sample data object based on dynamic bipolar modified probabilistic neural network (MPNN) |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2010102209511A CN101894185A (en) | 2010-06-29 | 2010-06-29 | Method for predicting life of small sample data object based on dynamic bipolar modified probabilistic neural network (MPNN) |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN101894185A true CN101894185A (en) | 2010-11-24 |
Family
ID=43103376
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN2010102209511A Pending CN101894185A (en) | 2010-06-29 | 2010-06-29 | Method for predicting life of small sample data object based on dynamic bipolar modified probabilistic neural network (MPNN) |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN101894185A (en) |
Cited By (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103018673A (en) * | 2012-11-19 | 2013-04-03 | 北京航空航天大学 | Method for predicating life of aerospace Ni-Cd storage battery based on improved dynamic wavelet neural network |
| CN103258115A (en) * | 2013-04-12 | 2013-08-21 | 中国航天标准化研究所 | Method for assessing dynamic reliability of lithium ion batteries on the basis of polynomial fitting and life distribution |
| CN103324834A (en) * | 2013-04-07 | 2013-09-25 | 北京航空航天大学 | Service life prediction method for electromechanical system and critical components under completely truncated data condition |
| CN107480395A (en) * | 2017-08-29 | 2017-12-15 | 燕山大学 | A kind of construction method and system of vehicle steering knuckle loading spectrum forecast model |
| CN107797067A (en) * | 2016-09-05 | 2018-03-13 | 北京航空航天大学 | Lithium ion battery life migration prediction method based on deep learning |
| CN108710905A (en) * | 2018-05-10 | 2018-10-26 | 华中科技大学 | One kind being based on the united spare part quantitative forecasting technique of multi-model and system |
| CN109492832A (en) * | 2018-12-24 | 2019-03-19 | 斑马网络技术有限公司 | Life-span prediction method, device, equipment and the storage medium of battery |
| CN109740802A (en) * | 2018-12-24 | 2019-05-10 | 斑马网络技术有限公司 | Discharged or defective battery prediction processing method, device, equipment and readable storage medium storing program for executing |
| CN109784779A (en) * | 2019-03-04 | 2019-05-21 | 广州杰赛科技股份有限公司 | Financial risk prediction technique, device and storage medium |
| CN111090051A (en) * | 2020-01-21 | 2020-05-01 | 北京空间飞行器总体设计部 | Method for automatically diagnosing discharge final voltage of cadmium-nickel storage battery for satellite |
| CN111794978A (en) * | 2020-07-23 | 2020-10-20 | 中国核动力研究设计院 | Safety injection pump operation life prediction method and system |
| CN113296009A (en) * | 2021-04-23 | 2021-08-24 | 重庆大学 | Method for predicting and recombining residual life of retired power lithium ion battery |
| CN114626474A (en) * | 2022-03-21 | 2022-06-14 | 厦门大学 | Vehicle power state detection method and system based on probabilistic neural network |
| CN114781278A (en) * | 2022-06-17 | 2022-07-22 | 天津理工大学 | Electromechanical equipment service life prediction method and system based on data driving |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP1542302A2 (en) * | 2003-11-18 | 2005-06-15 | Matsushita Electric Industrial Co., Ltd. | Fuel cell life predicting device and fuel cell system |
-
2010
- 2010-06-29 CN CN2010102209511A patent/CN101894185A/en active Pending
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP1542302A2 (en) * | 2003-11-18 | 2005-06-15 | Matsushita Electric Industrial Co., Ltd. | Fuel cell life predicting device and fuel cell system |
Non-Patent Citations (3)
| Title |
|---|
| 《2009 8th International Conference on reliability、Maintainability and safety》 20091231 Tao Lai-fa等 Prediction on Moonlet Power System Data Based On Modified Probability Neural Network , * |
| 《2010 International Conference on Electrical and Control Engineering》 20100627 Tao Lai-fa等 Life Prediction of Small Satellite Storage Battery Based on Dynamic Bi-Step MPNN Method , * |
| 《2010 International Conference on Electrical and Control Engineering》 20100627 Tao Lai-fa等 Life Prediction of Small Satellite Storage Battery Based on Dynamic Bi-step MPNN Method 第5305-5308页 , * |
Cited By (22)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103018673B (en) * | 2012-11-19 | 2015-01-21 | 北京航空航天大学 | Method for predicating life of aerospace Ni-Cd storage battery based on improved dynamic wavelet neural network |
| CN103018673A (en) * | 2012-11-19 | 2013-04-03 | 北京航空航天大学 | Method for predicating life of aerospace Ni-Cd storage battery based on improved dynamic wavelet neural network |
| CN103324834A (en) * | 2013-04-07 | 2013-09-25 | 北京航空航天大学 | Service life prediction method for electromechanical system and critical components under completely truncated data condition |
| CN103324834B (en) * | 2013-04-07 | 2016-08-17 | 北京航空航天大学 | Mechatronic Systems under a kind of completely cut off data qualification and critical component life-span prediction method thereof |
| CN103258115A (en) * | 2013-04-12 | 2013-08-21 | 中国航天标准化研究所 | Method for assessing dynamic reliability of lithium ion batteries on the basis of polynomial fitting and life distribution |
| CN103258115B (en) * | 2013-04-12 | 2016-06-08 | 中国航天标准化研究所 | Based on the lithium ion battery dynamic reliability appraisal procedure that fitting of a polynomial and life-span are distributed |
| CN107797067B (en) * | 2016-09-05 | 2019-09-27 | 北京航空航天大学 | Lithium ion battery life migration prediction method based on deep learning |
| CN107797067A (en) * | 2016-09-05 | 2018-03-13 | 北京航空航天大学 | Lithium ion battery life migration prediction method based on deep learning |
| CN107480395A (en) * | 2017-08-29 | 2017-12-15 | 燕山大学 | A kind of construction method and system of vehicle steering knuckle loading spectrum forecast model |
| CN107480395B (en) * | 2017-08-29 | 2019-12-13 | 燕山大学 | Method and system for constructing vehicle steering knuckle load spectrum prediction model |
| CN108710905B (en) * | 2018-05-10 | 2020-07-10 | 华中科技大学 | Spare part quantity prediction method and system based on multi-model combination |
| CN108710905A (en) * | 2018-05-10 | 2018-10-26 | 华中科技大学 | One kind being based on the united spare part quantitative forecasting technique of multi-model and system |
| CN109740802A (en) * | 2018-12-24 | 2019-05-10 | 斑马网络技术有限公司 | Discharged or defective battery prediction processing method, device, equipment and readable storage medium storing program for executing |
| CN109492832A (en) * | 2018-12-24 | 2019-03-19 | 斑马网络技术有限公司 | Life-span prediction method, device, equipment and the storage medium of battery |
| CN109784779A (en) * | 2019-03-04 | 2019-05-21 | 广州杰赛科技股份有限公司 | Financial risk prediction technique, device and storage medium |
| CN111090051A (en) * | 2020-01-21 | 2020-05-01 | 北京空间飞行器总体设计部 | Method for automatically diagnosing discharge final voltage of cadmium-nickel storage battery for satellite |
| CN111794978A (en) * | 2020-07-23 | 2020-10-20 | 中国核动力研究设计院 | Safety injection pump operation life prediction method and system |
| CN111794978B (en) * | 2020-07-23 | 2022-02-11 | 中国核动力研究设计院 | Safety injection pump operation life prediction method and system |
| CN113296009A (en) * | 2021-04-23 | 2021-08-24 | 重庆大学 | Method for predicting and recombining residual life of retired power lithium ion battery |
| CN114626474A (en) * | 2022-03-21 | 2022-06-14 | 厦门大学 | Vehicle power state detection method and system based on probabilistic neural network |
| CN114781278A (en) * | 2022-06-17 | 2022-07-22 | 天津理工大学 | Electromechanical equipment service life prediction method and system based on data driving |
| CN114781278B (en) * | 2022-06-17 | 2022-09-27 | 天津理工大学 | Electromechanical equipment service life prediction method and system based on data driving |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN101894185A (en) | Method for predicting life of small sample data object based on dynamic bipolar modified probabilistic neural network (MPNN) | |
| CN103018673B (en) | Method for predicating life of aerospace Ni-Cd storage battery based on improved dynamic wavelet neural network | |
| CN111584952B (en) | Method and system for online evaluation of electrochemical cells of energy storage power station | |
| Liang et al. | Analysis of factors affecting economic operation of electric vehicle charging station based on DEMATEL-ISM | |
| Hu et al. | A review of second-life lithium-ion batteries for stationary energy storage applications | |
| CN106327006A (en) | Comprehensive benefit analysis-based micro-power-grid optimal configuration method | |
| CN111738462B (en) | Fault first-aid repair active service early warning method for electric power metering device | |
| CN105701554A (en) | Electrical equipment energy efficiency chaos analysis method based on massive measurement data | |
| CN103488869A (en) | Wind power generation short-term load forecast method of least squares support vector machine | |
| CN109508826B (en) | Electric vehicle cluster schedulable capacity prediction method based on gradient lifting decision tree | |
| CN104376371B (en) | A kind of distribution based on topology is layered load forecasting method | |
| CN109307845A (en) | A kind of method and system for echelon using battery energy storage system progress health state evaluation | |
| CN104809658A (en) | Method for rapidly analyzing low-voltage distributing network area line loss | |
| CN108920609A (en) | Electric power experimental data method for digging based on multi dimensional analysis | |
| CN103983919A (en) | Battery life prediction method based on GM (1, N) gray model | |
| CN109759354A (en) | One kind retracting battery and shunts screening technique | |
| CN115121507B (en) | Retired power battery sorting method with low test cost | |
| CN114429050A (en) | Sorting method for gradient utilization of retired power batteries | |
| Manwell et al. | Improvements to the Hybrid2 battery model | |
| Ma et al. | Economic operation evaluation of active distribution network based on fuzzy Borda method | |
| CN114865668A (en) | Energy storage scheduling support evaluation method | |
| CN115545119A (en) | Method, system and application for identifying electricity consumption data | |
| CN112132424A (en) | Large-scale energy storage multi-attribute decision type selection method | |
| CN114723284A (en) | Reliability evaluation method for power distribution network comprising distributed power supply and electric automobile | |
| CN112734274B (en) | Low-carbon power grid operation leading influence factor mining and comprehensive evaluation method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C02 | Deemed withdrawal of patent application after publication (patent law 2001) | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20101124 |