CN103336908B - The method for predicting residual useful life of dull echo state network - Google Patents

Abstract

The method for predicting residual useful life of dull echo state network, belongs to failure prediction and system health management domain, the present invention be solve traditional E SN cannot the problem in Accurate Prediction life-span.The inventive method: step one, set up dull echo state network model M ONESN at random: step 2, carry out network dynamic training, obtain the monotone increasing of MONESN model or the output weights of monotone decreasing, these output weights are substituted in MONESN the MONESN obtained after training; Step 3, the residue operation week issue of the turbine engine in life-span to be predicted exported to the input block of the MONESN after step 2 training, the residual life being the turbine engine in life-span to be predicted that MONESN exports; Or the residue operation week issue of the lithium ion battery in life-span to be predicted exported to the input block of the MONESN after step 2 training, the residual life being the lithium ion battery in life-span to be predicted that MONESN exports.

Description

The method for predicting residual useful life of dull echo state network

Technical field

The present invention relates to the method for predicting residual useful life of dull echo state network, belong to failure prediction and system health management domain.

Background technology

Function approximation problem is a basic problem of neural network research, mainly relies on the nonlinearity approximation capability of neural network.Prove that feedforward network can approach the measurable function on any compact territory (Borel-territory).Recurrent neural network can be understood as the dynamic system of an opening, is described the mapping of recurrent neural network inside by the form of state space, as shown in Equation 1:

s _t+1=f(As _t+Bu _t+θ)statetransition

y _t=Cs _toutputequation(1)

Wherein, A, B, C are weight matrixs, and θ is deviation unit, are input variable u _tbiased.F is the activation function of neural network, normally S type function, such as tanh (thehyperbolictangent).Form due to the state transition equation of recurrent neural network makes the state S of subsequent time _t+1the whole input u before current time _t, u _t-1.... superposition, therefore recurrent neural network has short-term memory characteristic.RNN a kind of generally approaches device, and it can approach any open dynamic system by arbitrary accuracy.But because traditional Recursive Neural Network Structure is complicated, need to train whole weights; Learning method based on Gradient Descent causes the slow and height that assesses the cost of its speed of convergence.These two shortcomings of RNN limit its application.Along with the development of deposit pond computing technique reservoircomputing (RC), Jaeger professor proposes a kind of novel recurrent neural network in calendar year 2001, and be called echo state network (ESN), its network structure as shown in Figure 1.ESN takes into account the advantage of traditional neural network, overcomes the shortcoming of classic method simultaneously, has good Function approximation capabilities.ESN adopts dynamic deposit pond to replace the hidden layer of general neural network, and the neuron containing a large amount of partially connected in deposit pond, shows good memory characteristic; In addition, echo state network, in learning process, only needs training to export weights, thus simplifies the foundation of network.The advantage of ESN in Chaotic time series forecasting clearly, but is all the chaos time sequence for not Noise.But in the multi-Step Iterations forecasting process of the chaos time sequence of Noise, error can be accumulated by Single-step Prediction, and use linear regression method calculating output weights very easily to cause the morbid state solution of indirect problem, obtain the output weights of higher magnitude, document adopts the learning algorithm of regular terms, between the degree of accuracy of separating and flatness, make balance, be applied to the time series forecasting of Lorenz multistep, precision of prediction is higher than RBF model; We know, deposit pond scale, neuron behavior or other character all can have an impact to the performance of ESN, leakyintegrator neuron was used to replace the S type neuron of traditional E SN to carry out changing well afterwards, solve " figure8 " generationtask, and to classify Japanese vowel data with zero error rate; Proposed again the calculating of deposit pond and wavelet decomposition to combine afterwards to be used for time series forecasting, and adopted small echo original time series to be decomposed into the sequence of multiple comparatively easily prediction, after multiple ESN predicts respectively, the output sequence of ESN is reassembled into former sequence.The data set of ESTSP08 meeting is adopted to demonstrate the validity of method; Although ESN has the simple advantage of learning method, because the parameter of (1) (Thefreeparameters) ESN is very large to the performance impact of ESN, only have the method for expertise and cross validation that the parameter of ESN is set at present; (2) ESN only having single deposit pond is not omnipotent, and ESN can not be trained to as amultiplesuperimposedoscillator (MSO).For above two problems, uncoupled echo state network is proposed, use the embodiment of lateral inhibition method establishment two kinds of low complex degrees, be respectively DESN+ and lay in pond prediction (DESN+RP) and the maximum available information of DESN+ (DESN+MaxInfo).For predicting actual sea clutter data, compared with only having the ESN in single deposit pond, there is higher precision of prediction and robustness.But ESN is not still applied in the approximation problem of this class function with monotonic trend. in fact, this class function is very common in actual life.The purchasing power of such as common people height that is strong and weak and commodity price is proportional, and commodity price is higher, and purchasing power is more weak.Otherwise commodity price is lower, purchasing power is stronger; The body weight of the people also having us to know is larger, suffers from a heart complaint larger with the risk of chronic disease; Also have price and the location, house in house, room number becomes monotonic relationshi with floor space; Another typical example is fault diagnosis and fault prediction problem, relation between the residual life of Condition Monitoring Data and monitored object, the residual life of such as bearing and its two status monitoring parametric vibration frequencies and temperature are inversely proportional to, namely bear vibration frequency is larger, temperature is higher, and residual life is fewer etc.But traditional E SN is used in the approximation problem of the function with monotonic trend, there is the shortcomings such as counting yield is low, approximation capability is poor.

Summary of the invention

The present invention seeks in order to solve traditional E SN for there is the function of monotonic trend approximation problem in there is the shortcomings such as counting yield is low, approximation capability is poor, and then cannot the problem in Accurate Prediction life-span, provide a kind of method for predicting residual useful life of dull echo state network.

The method for predicting residual useful life of dull echo state network of the present invention, comprises the following steps:

Step one, set up dull echo state network model M ONESN at random:

Dull echo state network model M ONESN comprises input block u (n), interior processing unit x (n) and output unit y (n), and n is the moment of correspondence system state transfer; Input block gathers the analog physical parameter of turbine engine;

Output unit exports the residue operation week issue of turbine engine; Or output unit exports the residue operation week issue of lithium ion battery;

Set up N × L dimension input weight matrix W at random ⁱⁿ, N × N ties up inner connection weight value matrix W ₀, N × M ties up feedback weight matrix W ^back; L, M and N are positive integer;

The renewal equation of k moment interior processing unit is:

x(k)=f(W ⁱⁿu(k)+Wx(k-1)+W ^backy(k-1))

Wherein: W=α W ₁, W ₁=W ₀/ | λ _max|, wherein | λ _max| be the absolute value of the eigenvalue of maximum of W0, now W ₁spectral radius be 1;

Step 2, network dynamic training is carried out to the unbred MONESN that step one is set up, and then obtain the monotone increasing of MONESN model or the output weights of monotone decreasing, these output weights are substituted in MONESN the MONESN obtained after training;

Step 3, the residue operation week issue of the turbine engine in life-span to be predicted exported to the input block of the MONESN after step 2 training, the residual life being the turbine engine in life-span to be predicted that MONESN exports;

Or the residue operation week issue of the lithium ion battery in life-span to be predicted exported to the input block of the MONESN after step 2 training, the residual life being the lithium ion battery in life-span to be predicted that MONESN exports.

Advantage of the present invention: the present invention is directed to the MONESN model that the large shortcoming of approximate error that ESN exists when the function approaching monotonic trend proposes, and this model is used for two examples in predicting residual useful life field: the predicting residual useful life problem of turbine engine and battery, owing to adding the constraint of monotonicity in the training process, reduce the scope of optimization, improve and approach efficiency; From approximation capability aspect, for the more obvious data of monotonicity, the advantage of MONESN is more obvious.In addition, data have this feature of monotonicity than being easier to discovery, but other prioris existed in data as consistance etc. be do not allow detectable, the information of image watermarking can be found for different data, add in the training process of ESN so that the approximation capability of better lifting function in the mode of constraint equally.

Accompanying drawing explanation

Fig. 1 is traditional echo state network (ESN) model in background technology; LInputUnit is that L ties up input block, and NInternalUnits is that N ties up interior processing unit, and MOutputUnits is that M ties up output unit.

Fig. 2 is the process flow diagram of the method for predicting residual useful life of dull echo state network of the present invention;

Fig. 3 to Fig. 5 is the graph of a relation of turbine engine inputoutput data;

Fig. 6 is MONESN and ESN training stage error comparison diagram;

Fig. 7 is MONESN and ESN test phase error comparison diagram;

Deposit pond scale figure when Fig. 8 is the best performance of MONESN and ESN;

Fig. 9 is the first cell set capacity degenerated curve; Figure 10 is the 3rd cell set capacity degenerated curve;

Figure 11 is the B18 lithium ion battery residual life prediction curve based on ESN

Figure 12 is the B18 lithium ion battery residual life prediction curve based on MONESN.

Embodiment

Embodiment one: present embodiment is described below in conjunction with Fig. 2, the method for predicting residual useful life of dull echo state network described in present embodiment, step one, set up dull echo state network model M ONESN at random:

Dull echo state network model M ONESN comprises input block u (n), interior processing unit x (n) and output unit y (n), and n is the moment of correspondence system state transfer; Input block gathers the analog physical parameter of turbine engine, comprises the speed parameter of temperature, pressure and system; Or input block gathers the discharge and recharge data of lithium ion battery;

U (n) represents the value in input block n moment, and x (n) represents the value in interior processing unit n moment, and y (n) represents the value in output unit n moment.

Output unit exports the residue operation week issue of turbine engine; Or output unit exports the residue operation week issue of lithium ion battery;

Set up N × L dimension input weight matrix W at random ⁱⁿ, N × N ties up inner connection weight value matrix W ₀, N × M ties up feedback weight matrix W ^back; L, M and N are positive integer;

The renewal equation of k moment interior processing unit is:

x(k)=f(W ⁱⁿu(k)+Wx(k-1)+W ^backy(k-1))

Wherein: W=α W ₁, W ₁=W ₀/ | λ _max|, wherein | λ _max| be W ₀the absolute value of eigenvalue of maximum, now W ₁spectral radius be 1;

Step 2, network dynamic training is carried out to the unbred MONESN that step one is set up, and then obtain the monotone increasing of MONESN model or the output weights of monotone decreasing, these output weights are substituted in MONESN the MONESN obtained after training;

Step 3, the residue operation week issue of the turbine engine in life-span to be predicted exported to the input block of the MONESN after step 2 training, the residual life being the turbine engine in life-span to be predicted that MONESN exports;

Or the residue operation week issue of the lithium ion battery in life-span to be predicted exported to the input block of the MONESN after step 2 training, the residual life being the lithium ion battery in life-span to be predicted that MONESN exports.

Embodiment two: present embodiment is described further embodiment one, set up dull echo state network model M ONESN described in step one at random and comprise input block u (n), interior processing unit x (n) and output unit y (n)

u(n)=(u ₁(n),u ₁(n),...,u _j(n),...,u _L(n)),j=1,2,...,L；

x(n)=(x ₁(n),x ₂(n),...,x _s(n),...,x _N(n)),s=1,2,...,N；

y(n)=(y ₁(n),y ₂(n),...,y _i(n),...,y _M(n)),i=1,2,...,M；

N is the moment of correspondence system state transfer, and L, M and N are positive integer;

The process setting up MONESN is:

Step sets up N × L dimension input weight matrix W one by one, at random ⁱⁿ, N × N ties up inner connection weight value matrix W ₀feedback weight matrix W is tieed up with N × M ^back;

Step one two, according to formula:

W ₁=W ₀/|λ _max|

Obtain W ₁, wherein | λ _max| be W ₀the absolute value of eigenvalue of maximum, now W ₁spectral radius be 1;

Step one three, according to formula:

W=αW ₁

Obtain the renewal equation of k moment interior processing unit

X (k)=f (W ⁱⁿu (k)+Wx (k-1)+W ^backy (k-1)) parameter W;

Set up unbred MONESN.

Embodiment three: present embodiment is described further embodiment one, the process that the unbred MONESN set up step one described in step 2 carries out network dynamic training is:

Step 2 one, initialize internal processing unit x (0)=0, output unit y (0)=0;

Step 2 two, list entries u _j(n) and true output sample sequences y _in () drives MONESN, according to the interior processing unit renewal equation of MONESN

x(n)=f(W ⁱⁿu(n)+Wx(n-1)+W ^backy(n-1))

With

y(n)=W ^out(u(n),x(n))

Obtain the intrinsic nerve unit state in each moment;

The input in each moment and internal state are deposited T × (L+N) with the form of row vector to be tieed up in internal state Matrix C; And the true output in corresponding moment is deposited T × M with the form of row vector is in output matrix d;

Transposition composition L × (L+N) that step 2 three, structure constraint inequality: L tie up unit matrix and L × N dimension input weights ties up matrix A; B is that null matrix is tieed up in L × 1.

Embodiment four: present embodiment is described further embodiment one, the process exporting the output weights of monotone increasing or monotone decreasing described in step 3 is:

Press $\begin{array}{c}\underset{{\left(W^{\mathrm{out}}\right)}^T}{\min }\frac{1}{2}{\left(\mathrm{norm}(C\×{\left(W^{\mathrm{out}}\right)}^T-d)\right)}^2\\ s.t.A\×{\left(W^{\mathrm{out}}\right)}^T\≤b\end{array}$ Export the output weights transposition (W with monotone decreasing relation ^out) ^t, after transposition, export the output weights W with monotone decreasing relation ^out;

Press $\begin{array}{c}\underset{{\left(W^{\mathrm{out}}\right)}^T}{\min }\frac{1}{2}{\left(\mathrm{norm}(C\×{\left(W^{\mathrm{out}}\right)}^T-d)\right)}^2\\ s.t.A\×{\left(W^{\mathrm{out}}\right)}^T\≤b\end{array}$ Export the output weights transposition (W with monotone increasing relation ^out) ^t, after transposition, export the output weights W with monotone increasing relation ^out.

As shown in Figure 1, comprise L to tie up input variable, N be internal state variable and M is output variable to the structure of ESN, and for convenience's sake, we suppose W ^back=0, input/output relation meets following relational expression:

In order to ensure the output variable y of ESN _i(i ∈ [1 ..., M]) and input variable u _j(j ∈ [1 ..., L]) there is monotonically increasing relation, namely to above-mentioned relation formula to u _jask local derviation, can following formula be obtained:

$\frac{{\∂y}_i}{{\∂u}_j}=w_{\mathrm{ij}}^{\mathrm{out}}+\underset{t=1}{\overset{N}{\mathrm{\Σ}}}{w}_{j*(L+t)}^{\mathrm{out}}(1-{\mathrm{\θ}}_t^2)w_{\mathrm{tj}}^{\mathrm{in}}>0$

Known: the derivative of tanh is necessarily greater than zero, output variable y _iwith input variable u _jthe adequate condition of monotonically increasing relation is kept to be shown below:

$\frac{{\∂y}_i}{{\∂u}_j}=w_{\mathrm{ij}}^{\mathrm{out}}+\underset{t=1}{\overset{N}{\mathrm{\Σ}}}{w}_{j*(L+t)}^{\mathrm{out}}{w}_{\mathrm{tj}}^{\mathrm{in}}>0,\∀i\·j$

Comprising two: one in above formula is connect to export y _iwith input u _joutput weight w _ij ^out, another connects to export y _iwith internal state x _t(t ∈ [1 ..., N]) output weight w _{j (L+t)} ^outu is inputted with being connected _jwith internal state x _tproduct and, if the input weights of ESN and export weights meet above formula, can output variable y be ensured _irelative to input variable u _jit is monotonically increasing.

Proving by the same methods, ensures output variable y _irelative to input variable u _jbe the adequate condition of monotone decreasing be that two sums are less than zero, be shown below:

$\frac{{\&PartialD;y}_i}{{\&PartialD;u}_j}=w_{\mathrm{ij}}^{\mathrm{out}}+\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}{w}_{j*(L+k)}^{\mathrm{out}}{w}_{\mathrm{ki}}^{\mathrm{in}}<0,\&ForAll;i,j$

Can know thus: if want to remove to approach the function with monotonic trend with ESN, in the training process of ESN, so add the output variable y that constraint just can ensure ESN _irelative to ESN input variable u _jthe relation that there is monotone increasing or successively decrease.

Embodiment five: present embodiment is described below in conjunction with Fig. 3 to Fig. 8, present embodiment provides a specific embodiment predicted the residual life of turbine engine.

Present embodiment adopts root-mean-square error RMSE and R ²two kinds of error calculation method evaluate MONESN model has the function of dull input/output relation approximation capability to a class.Suppose given function for function average.The output valve of MONESN is f ' (x), and wherein x is input variable, and data length is n.

In approximation problem, often adopt root-mean-square error (RootMeanSquaredError, RMSE) as the evaluation index of approximation capability, be shown below:

$\mathrm{RMSE}\left(x\right)=\sqrt{\frac{\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{(f\left(x_i\right)-f^{\&prime;}\left(x_i\right))}^2}{n}}$

Another kind of evaluation method adopts R ²the overall fit effect of evaluation function, when the fitting effect of model non-constant time, model output valve can be greater than the error sum of squares of the average of model output valve and actual value in the quadratic sum of the error of actual value, i.e. R ²negative value may be there is.R ²evaluation function is shown below:

$R^2=1-\frac{{\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}(f\left(x_i\right)-f^{\&prime;}\left(x_i\right))}^2}{{\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}(f\left(x_i\right)-\stackrel{\&OverBar;}{f}\left(x_i\right))}^2}.$

Present embodiment adopts the validity of turbine engine degeneration emulated data verification method, and data are the public data collection coming from NASA's Ames predicted data warehouse.CMAPSS can the state of simulated engine model under different operating conditionss, and by changing 13 health parameters on CMAPSS, user can the impact of any one fault and degradating trend in simulated engine 5 rotary parts.Simultaneously CMAPSS provides 58 different output variables, and in the simulated experiment that Saxenaetal. carries out, 21 output variable analog sensor measuring systems in collecting 58, comprise temperature, the speed of pressure and system difference.Also have collected characterizing engine operation condition 3 variablees in addition. this 24 dimension data features engine and runs until the state trajectory of the overall process lost efficacy.

NASA provides 4 independently 4 group data sets of obtaining of emulation experiment.Every group data set comprises training and testing subdata collection and residual life corresponding to the sample concentrated of test subdata, training set comprises multiple training engine sample and runs until the total data of inefficacy, and test engine is run until lost efficacy, but data centralization only gives to be run until the data of certain specific period before losing efficacy, the life-span of engine is defined as total periodicity of engine operation, so the residual life of test sample book is total periodicity that engine runs deduct the specific period number provided in test set.The difference of every group data set is: training and testing number of samples, different types of operating conditions and fault mode.Specifically arranging in turbine engine degeneration emulation experiment process is as shown in table 1, comprise the residual life of 100 training engine samples and 100 test data set samples and 100 test sample books in first group data set, have a kind of operating conditions and a kind of failure mode.Other data set contents are similar, no longer introduce at this.

The similarities and differences of table 14 group data set

By the analysis to data, because data set 2 and 4 is containing 6 kinds of operator schemes, namely before, the data of three-dimensional reflection engine operating condition have impact to residual life, cause there is not monotonic relationshi, so do not consider this two data sets herein between 21 of these two groups of data dimension Sensor monitoring data and residual life.Remaining data set 1 and 3 all contains a kind of operator scheme, and front three-dimensional data is little on residual life impact, there is monotonic relationshi between 21 dimension Sensor monitoring data and residual life, adopts the validity of the 1st data set checking put forward the methods herein.The training and testing subdata collection comprised in 1st group data set is minimum/large running length and average length as shown in table 2.

Table 2 is trained and the length of test cell is added up

Engine often runs one-period residual life and subtracts 1, so to be the relation tieed up between sensor data measured and residual life with the .21 of slope-1 monotone decreasing as shown in table 3 the life-span of engine, and wherein 1,5,6,10,16,18, the data that No. 19 sensors record are constants, and remain unchanged between exporting, 7,12,20, the general trend of the data that No. 21 sensors record declines, because residual life is monotone decreasing, is monotonically increasing relation between known input and output.In like manner 2,3,4,8,9,11,13,14, the data general trend that 15, No. 17 sensors record rises, so keep the relation of monotone decreasing between input and output.Fig. 3 is to Figure 5 shows that 1,7, graph of a relation between No. 2 sensors and output, the residual life of what wherein the longitudinal axis represented is turbine engine, transverse axis is the data that three sensors record respectively, the gap of the numerical value recorded due to different sensors is comparatively large, so first carry out normalization to the data that each sensor except steady state value records, the relation of input and output can not be affected by the linear module of getting when monitoring.Normalization is calculated as follows shown in formula:

$z_i=\frac{x_i-\underset{1\&le;j\&le;n}{\min }{x}_j}{\underset{1\&le;j\&le;n}{\max }{x}_j-\underset{1\&le;j\&le;n}{\min }{x}_j}i=\mathrm{1,2},...,n$

The relation existed between table 3 input and output data

Add up the Monitoring Data of each turbine engine unit, select that wherein there are 22 unit that output data have more obvious monotonic trend: 17,20,24,27,32,34,37,40,41,42,49,53,56,58,61,64,76,77,82,90,98,100. as experimental data.First the input of MONESN, output unit is determined.Output unit is the residual life of each turbine engine, successively decreases for-1 with slope.Input block is Condition Monitoring Data, because said Condition Monitoring Data has three kinds above: constant, increases progressively and successively decreases.Can be judged by intuition, the data remained unchanged to whole prediction modeling process should be do not have effective, so do not consider this kind of data.The so remaining Monitoring Data increasing progressively and successively decrease, because the unit that sensor 8,14 and 17 is having increases progressively, some unit successively decrease, so using 11 remaining dimension data as input data.

After input-output unit is determined, use method training ESN and MONESN of cross validation.The maximum method of current use is cross validation, as shown in table 4, the scope of wherein laying in pond scale N is [10,80], and step-length is 5. spectral radius sr, the span [0.01 of input block yardstick IS and input block displacement IF, 1], step-length be 0.05. in order to contrast with the experimental result of traditional E SN, use cross validation to select deposit pond parameter to ESN equally, difference is the deposit pond scale span of ESN is [10,100].

Table 4MONESN and ESN cross validation optimum configurations

Thus, according to the training step of MONESN of 3.2 joints, obtain optimum output weights, the error result of experiment as shown in Figure 6, is greater than ESN and R at the RMSE of training stage MONESN ²generally be less than ESN, that is the performance ESN of training stage is better than MONESN.Shown in test phase error as shown in Figure 7, result is divided into two kinds of situations: the first situation is that the performance of MONESN is better than ESN, can prove the validity of method in this paper.The second situation is that ESN is better than or is more or less the same with the performance of MONESN.Wherein, the calculating of error calculates the mean value of trying to achieve 100 times.Shown in Fig. 8 is the value of the deposit pond scale of ESN and MONESN when obtaining best Approximation effect, we can see intuitively, time MONESN wants to reach the Approximation effect identical with ESN or even Approximation effect more better than ESN, the deposit pond intrinsic nerve unit needed will be significantly less than ESN, can reduce the calculated amount of model. and this is an advantage of MONESN.

Embodiment six: present embodiment is described below in conjunction with Fig. 9 to Figure 12, present embodiment provides a specific embodiment predicted the residual life of lithium ion battery.

Present embodiment adopts the lithium ion battery public data collection that provides of NASAAMESPCoE research centre Idaho National Laboratory of associating USDOE. and experiment is charged to battery under room temperature (25 DEG C), discharges and impedance measurement.NASA provides 3 groups of 3 group data sets obtained under different experimental conditions.First group is 25 to No. 28 batteries, and second group is 5,6 of 25 to No. 44 batteries and the 3rd group, No. 7 and No. 18 batteries.The data of battery provide with the form of Array for structural body, comprises type, ambienttemperature, time and Monitoring Data structure, has detailed introduction on concrete Monitoring Data structure website.

The capacity of battery declines gradually along with charging-discharging cycle, and when the charging capacity of battery drops to the 70%-80% of rated capacity, the serviceable life of battery terminates.The Cell Experimentation An of NASA adopts the service life cycle of the degenerate state characterizing battery of the capacity of battery.The capacity of battery can not directly be measured, but owing to there is the linear dependence of height between battery capacity and internal driving, so the internal resistance of cell can be measured according to battery impedance, then estimate battery capacity by data fitting method, battery capacity data is stored in the Monitoring Data group of electric discharge type.

It is the comparison diagram of first group and the 3rd cell set capacity degenerated curve shown in Fig. 9 and Figure 10, find that the data of the 3rd Battery pack present obvious degenerative character, there is obvious monotonic nature, so select the validity of the 3rd group of data verification method herein, introduce the experimentation of No. 18 batteries herein.The capacity data of No. 18 batteries totally 132, adopts Single-step Prediction herein, and wherein 66 data are as training data, is used for Modling model, and 66 data, as test data, are used for the performance of verification method.The method of same employing cross validation chooses the parameter of MONESN, optimum MONESN optimum configurations is as shown in table 5, predict the outcome as is illustrated by figs. 11 and 12, in the training stage, the performance of ESN and MONESN is all fine, difference is embodied in test phase, two kinds of errors of MONESN and ESN as shown in Table 6 relatively, the R of ESN ²<0, can illustrate that the spatial approximation effect of ESN is bad, and MONESN can follow the tracks of the residual life of battery very accurately.

Table 5ESN optimum configurations

Two kinds of error contrasts of table 6#18 battery two kinds of models

Owing to adding the constraint of monotonicity in the training process, reduce the scope of optimization, improve and approach efficiency.

Claims (5)

1. the method for predicting residual useful life of dull echo state network, is characterized in that, the method comprises the following steps:

Step one, set up dull echo state network model M ONESN at random:

Dull echo state network model M ONESN comprises input block u (n), interior processing unit x (n) and output unit y (n), and n is the moment of correspondence system state transfer;

Input block gathers the analog physical parameter of turbine engine; Output unit exports the residue operation week issue of turbine engine;

Or input block gathers the discharge and recharge data of lithium ion battery; Output unit exports the residue operation week issue of lithium ion battery;

Set up N × L dimension input weight matrix W at random ⁱⁿ, N × N ties up inner connection weight value matrix W ₀, N × M ties up feedback weight matrix W ^back; L, M and N are positive integer;

The renewal equation of k moment interior processing unit is:

x(k)＝f(W ⁱⁿu(k)+Wx(k-1)+W ^backy(k-1))

Wherein: W=α W ₁, W ₁=W ₀/ | λ _max|, wherein | λ _max| be W ₀the absolute value of eigenvalue of maximum, now W ₁spectral radius be 1;

Step 2, network dynamic training is carried out to the unbred MONESN that step one is set up, and then obtain the monotone increasing of MONESN model or the output weights of monotone decreasing, these output weights are substituted in MONESN the MONESN obtained after training;

Step 3, the residue operation week issue of the turbine engine in life-span to be predicted exported to the input block of the MONESN after step 2 training, the residual life being the turbine engine in life-span to be predicted that MONESN exports;

Or the residue operation week issue of the lithium ion battery in life-span to be predicted exported to the input block of the MONESN after step 2 training, the residual life being the lithium ion battery in life-span to be predicted that MONESN exports.

2. the method for predicting residual useful life of dull echo state network according to claim 1, it is characterized in that, set up dull echo state network model M ONESN described in step one at random and comprise input block u (n), interior processing unit x (n) and output unit y (n)

u(n)＝(u ₁(n),u ₁(n),...,u _j(n),...,u _L(n)),j＝1,2,...,L；

x(n)＝(x ₁(n),x ₂(n),...,x _s(n),...,x _N(n)),s＝1,2,...,N；

y(n)＝(y ₁(n),y ₂(n),...,y _i(n),...,y _M(n)),i＝1,2,...,M；

N is the moment of correspondence system state transfer, and L, M and N are positive integer;

The process setting up MONESN is:

Step sets up N × L dimension input weight matrix W one by one, at random ⁱⁿ, N × N ties up inner connection weight value matrix W ₀feedback weight matrix W is tieed up with N × M ^back;

Step one two, according to formula:

W ₁＝W ₀/|λ _max|

Obtain W ₁, wherein | λ _max| be W ₀the absolute value of eigenvalue of maximum, now W ₁spectral radius be 1;

Step one three, according to formula:

W＝αW ₁

Obtain the renewal equation of k moment interior processing unit

X (k)=f (W ⁱⁿu (k)+Wx (k-1)+W ^backy (k-1)) parameter W;

Set up unbred MONESN.

3. the method for predicting residual useful life of dull echo state network according to claim 2, is characterized in that, the process that the unbred MONESN set up step one described in step 2 carries out network dynamic training is:

Step 2 one, initialize internal processing unit x (0)=0, output unit y (0)=0;

Step 2 two, list entries u _j(n) and true output sample sequences y _in () drives MONESN, according to the interior processing unit renewal equation of MONESN

x(n)＝f(W ⁱⁿu(n)+Wx(n-1)+W ^backy(n-1))

With

y(n)＝W ^out(u(n),x(n))

Obtain the intrinsic nerve unit state in each moment;

The input in each moment and internal state are deposited T × (L+N) with the form of row vector to be tieed up in internal state Matrix C; And the true output in corresponding moment is deposited T × M with the form of row vector is in output matrix d;

Transposition composition L × (L+N) that step 2 three, structure constraint inequality: L tie up unit matrix and L × N dimension input weights ties up matrix A; B is that null matrix is tieed up in L × 1.

4. the method for predicting residual useful life of dull echo state network according to claim 3, it is characterized in that, the process exporting the output weights of monotone increasing or monotone decreasing described in step 3 is:

Press $\begin{array}{cc}\underset{{\left(W^{out}\right)}^T}{\min }& \frac{1}{2}{\left(norm\right(C\&times;{\left(W^{out}\right)}^T-d\left)\right)}^2\\ s.t.& A\&times;{\left(W^{out}\right)}^T\&le;b\end{array}$ Export the output weights transposition (W with monotone decreasing relation ^out) ^t, after transposition, export the output weights W with monotone decreasing relation ^out;

Press $\begin{array}{cc}\underset{{\left(W^{out}\right)}^T}{\min }& \frac{1}{2}{\left(norm\right(C\&times;{\left(W^{out}\right)}^T-d\left)\right)}^2\\ s.t.& A\&times;{\left(W^{out}\right)}^T\&GreaterEqual;b\end{array}$ Export the output weights transposition (W with monotone increasing relation ^out) ^t, after transposition, export the output weights W with monotone increasing relation ^out.

5. the method for predicting residual useful life of dull echo state network according to claim 1, is characterized in that, the analog physical parameter that input block gathers turbine engine comprises the speed parameter of temperature, pressure and system; Or input block gathers the discharge and recharge data of lithium ion battery.