CN104978459A - IGBT remaining useful life prediction method - Google Patents

Abstract

The present invention discloses an IGBT remaining useful life prediction method. By a phase space framing reconstruction technology, on the basis of reconstructing a phase space according to a differential entropy rate, the reconstructed phase space is subjected to Volterra-series input signal vector arrangement; correlation between input data and target output is considered, the optimal selection of each frame of input data is carried out, and the IGBT remaining useful life prediction method adopts a forward-backward algorithm and a least angle regression algorithm which are mature currently to use the optimal input data in input vectors as inputs of a model; a multiresponse sparse regression algorithm and a one-by-one extraction method are added on the basis of an original ELM model to cut off hidden layer nodes which are useless or have little effects, and three mixed neuronal activation functions are used, so that an established network has higher robustness and generalization. According to the present invention, difference of different inputs on a prediction model is sufficiently considered; the prediction model capable of dynamically updating each set of input data by an adaptive algorithm is designed; and prediction accuracy is greatly improved.

Description

A kind of IGBT method for predicting residual useful life

Technical field

The invention belongs to Novel power semiconductor device reliability analysis technical field, more specifically say, a kind of IGBT method for predicting residual useful life.

Background technology

Insulated gate bipolar transistor (IGBT, Insulated Gate Bipolar Transistor) as device for power switching, it has many advantages such as current carrying density is large, saturation pressure reduction, has been widely used at present in many key areas such as generation of electricity by new energy, high voltage power transmission.

As the core of system, the reliability effect of IGBT the operation stability of whole system equipment, this makes the residual life (RUL to IGBT, Remaining Useful Life) research of Forecasting Methodology becomes very necessary, it has some significance following: (1) is the important channel obtaining IGBT reliability information, may further be to realize system on-line monitoring and health control provides foundation; (2) contribute to impelling manufacturer to carry out process modification (introduce new material and improve encapsulation technology) to IGBT; (3) be conducive to designing accelerated aging test better to obtain aging data more accurately; (4) can condition maintenarnce be realized, make terminal user obtain the more life information of IGBT to reduce the input to system maintenance.

The research of the existing method for predicting residual useful life to IGBT is mainly divided into two large classes: the forecasting techniques of physically based deformation model-driven and the forecasting techniques based on data-driven.Although the forecasting techniques of physically based deformation model-driven can state the residual life information of IGBT from the angle of material, but need there is enough dark understanding to the intrinsic propesties of device manufactured materials and industrial manufacturing process, but the physical model obtained often lacks enough precision, nonlinear relationship between the easy lost objects parameter of process of modeling, causes error to increase.On the other hand, constructed physical model often with device concrete model closely related, the product category that on this and market, IGBT rapidly increases produces contradiction, and thus these class methods must bring the hysteresis quality in Forecasting Methodology.And be from the mapping relations between the historical aging data learning input and output of IGBT based on the forecasting techniques of data-driven, then set up non-linear, the nontransparent and non-model for special object in inside, in order to the remaining lifetime value of calculating device.If IGBT forecast model accurately can be set up, greatly will improve precision of prediction, thus reduce predicated error further.But also fewer based on the IGBT method for predicting residual useful life of data-driven at present, and the accuracy of prediction need to improve.

Summary of the invention

The object of the invention is to overcome the deficiencies in the prior art, a kind of IGBT method for predicting residual useful life is provided, to improve the precision of prediction of residual life.

For achieving the above object, IGBT method for predicting residual useful life of the present invention, is characterized in that, comprise the following steps:

(1), by accelerated life test, obtain IGBT collection emitter-base bandgap grading saturation voltage drop aging data time series, obtain data set DS={x (1), x (2) ..., x (D) }, wherein D is the data amount check obtained;

(2), first, for data set DS, best Embedded dimensions d and delay time T is determined according to the method for differential entropy rate; Then, utilize the function windowize in Matlab to carry out framing reconstruct to data set DS data set, and each frame data are mapped to d dimensional feature space respectively, thus obtain n (n=D-d τ) individual Frame { (X _t,y _t), wherein t=1,2 ..., n, to export as the input data of training pattern and target, wherein:

X _t＝{1，x(t)，x(t+τ)，…，x(t+(d-1)τ)，

x ²(t)，x(t)x(t+τ)，…x(t)x(t+(d-1)τ)，

x ²(t+τ)，x(t+τ)x(t+2τ)，…，x(t+τ)x(t+(d-1)τ)，

x ²(t+2τ)，x(t+2τ)x(t+3τ)，…，x(t+2τ)x(t+(d-1)τ)

，…，

x ²(t+(d-2)τ)，x(t+(d-2)x(t+(d-1)τ)

x ²(t+(d-1)τ)}；

Y _t＝x(t+dτ)

(3) Forward-backward algorithm (FB, is utilized, Forward-Backward algorithm) or minimum angle regression algorithm (LARS, Least Angle Regression algorithm) every frame input data X that step (2) is obtained _tdo optimal selection, select and export Y with corresponding target _tm the input data that correlativity is higher, the every frame input data after note selection are:

X _t'={ 1, x (t+c ₁τ), x (t+c ₂τ) ..., x (t+c _bτ) ..., x (t+c _eτ) x (t+c _fτ) }; c _b, c _e, c _f∈ 0,1 ..., d-1}, and c _e≤ c _f;

Then every frame input data X _t' dimension size (element number) be m, can again be expressed as:

X _t′＝[x _t,1,x _t,2,…,x _t,m] ^T(t＝1,2,…,n)；

(4), extreme learning machine (ELM) model of initialization one containing N (N < D) individual hidden layer node; The data of input layer are the input data X that step (3) obtains _t' (t=1,2 ..., n);

At the input weights matrix of coefficients of interval [-1,1] interior random initializtion ELM with threshold coefficient matrix wherein s=1,2 ..., N; I=1,2 ..., m;

According to initialized input weights matrix of coefficients with threshold coefficient matrix calculate hidden layer output matrix wherein network hidden layer output vector for:

${\stackrel{\&OverBar;}{h}}_t={\&lsqb;g({\mathrm{\&Sigma;}}_{i=1}^m{\stackrel{\&OverBar;}{w}}_{1,i}{x}_{t,i}-{\stackrel{\&OverBar;}{\mathrm{\&theta;}}}_1),g({\mathrm{\&Sigma;}}_{i=1}^m{\stackrel{\&OverBar;}{w}}_{2,i}{x}_{t,i}-{\stackrel{\&OverBar;}{\mathrm{\&theta;}}}_2),...,g({\mathrm{\&Sigma;}}_{i=1}^m{\stackrel{\&OverBar;}{w}}_{N,i}{x}_{t,i}-{\stackrel{\&OverBar;}{\mathrm{\&theta;}}}_N)\&rsqb;}^T,$ G represents excitation function;

(5), according to the vector that step (4) obtains and the corresponding target obtained in step (2) exports Y _t, utilize multiple response sparse regression algorithm (MRSR, Multiresponse Sparse Regression) to vector element and hidden layer node rearrange, note sequence after hidden layer node be:

wherein, subscript 1≤j _i≤ N is the hidden layer node sequence number before sequence, and subscript 1≤i≤N is the hidden layer node precedence sequence number after sequence; According to the order sequenced, l hidden layer node, wherein l < N before extracting;

(6), according to step (5) extract input weights coefficient corresponding to hidden layer node and threshold coefficient, input weights coefficient and threshold coefficient matrix are updated to w=(w respectively _s,i) _{l × m}with θ=θ _ss=1,2 ..., l; I=1,2 ..., m, and to after extraction (beta pruning), recalculate hidden layer output matrix H=[h ₁, h ₂..., h _j..., h _n], wherein vector $h_j={\left[\begin{array}{ccc}g({\mathrm{\&Sigma;}}_{i=1}^m{w}_{1,i}{x}_{j,i}-{\mathrm{\&theta;}}_1)& ...& g({\mathrm{\&Sigma;}}_{i=1}^m{w}_{l,i}{x}_{j,i}-{\mathrm{\&theta;}}_l)\end{array}\right]}^T;$

(7), solve beta pruning after network export weights and predicting the outcome, its process is as follows:

7.1), predict D+1 data, namely time predicted data x (D+1), the original input data frame according to step (2) and (3) known forecast model is X _t, now, t=D+1-d τ; When after input selection, note selects rear input data frame to be X _t', t=D+1-d τ, then can calculate the network hidden layer output vector in this moment and t=D+1-d τ moment $h_t={\left[\begin{array}{ccc}g\left({\mathrm{\&Sigma;}}_{i=1}^m{w}_{1,i}{x}_{t,i}-{\mathrm{\&theta;}}_1\right)& ...& g\left({\mathrm{\&Sigma;}}_{i=1}^m{w}_{l,i}{x}_{t,i}-{\mathrm{\&theta;}}_l\right)\end{array}\right]}^T;$

7.2), computational grid hidden layer output vector h _tthe hidden layer output matrix H=[h obtained with step (6) ₁, h ₂..., h _j..., h _n] in each vectorial h _j, j=1,2 ..., the Euclidean distance S of n;

7.3), to network hidden layer output vector h in Euclidean distance S _twith each vectorial h of hidden layer output matrix H _j, j=1,2 ..., the distance value of n sorts, and finds out closest to network hidden layer output vector h from hidden layer output matrix H _tz, z>=l the new matrix of vector composition and the target of correspondence exports ${\hat{Y}}_{z\&times;1}={\&lsqb;{\hat{Y}}_1,{\hat{Y}}_2,...,{\hat{Y}}_z\&rsqb;}^T,$ Thus calculating exports weights accordingly $\mathrm{\&gamma;}={\&lsqb;{\mathrm{\&gamma;}}_1,{\mathrm{\&gamma;}}_2,...,{\mathrm{\&gamma;}}_l\&rsqb;}^T={{\hat{H}}_{z\&times;l}}^{+}{\hat{Y}}_{z\&times;l}={\left({\hat{H}}^T\hat{H}\right)}^{-1}{\hat{H}}^T\hat{Y};$

The t=D+1-d τ moment is obtained, the predicted value of forecast model by these output weights:

(8), the predicted value that obtains of determining step (7) whether reach the performance failure threshold value (APT, Acceptable Performance Threshold) of IGBT, if reached, then this predicted value position be D+1 as critical point record, stop prediction, if do not reached, then by predicted value join IGBT collection emitter-base bandgap grading saturation voltage drop aging data time series, obtain data set DS={x (2), x (3) ..., x (D+1) }, then obtain predicted value according to the method for step (1) ~ (7) and judge whether the performance failure threshold value reaching IGBT, if no, then continue to join IGBT collection emitter-base bandgap grading saturation voltage drop aging data time series, obtain data set DS={x (3), x (4),, x (D+2) }, repeat like this, until reach, and using this predicted value position D+W as critical point record, wherein, W is predicted value reach the prediction number of times of the performance failure threshold value of IGBT.

The object of the present invention is achieved like this.

IGBT method for predicting residual useful life of the present invention, by phase space framing reconfiguration technique, on the basis of differential entropy rate phase space reconstruction, in order to pursue higher precision of prediction, change the prediction being predicted as adjacent data difference to IGBT collection emitter-base bandgap grading saturation voltage drop data, to the phase space reconstructed with the input signal vector arrangement of Volterra progression; Consider the correlativity that input data and target export, namely some input is to predicting inoperative or plaing a part very little, carry out the optimal selection of each frame input data, the present invention adopts the Forward-backward algorithm of current comparative maturity and minimum angle regression algorithm to select preferably to input the input that data are used as model in input vector; On the basis of original ELM model, add multiple response sparse regression algorithm and one by one extraction method crop useless or act on little hidden layer node, and use three kinds of neuron activation functions of mixing, thus set up network is made to have more robustness and generalization; The present invention considers the otherness of different inputs to forecast model fully, devises the forecast model that a kind of KALE of utilization adaptive algorithm upgrades each group input data dynamically, greatly improves the precision of prediction.

Therefore, the present invention draws the Accurate Model of Volterra progression and the optimum beta pruning extreme learning machine (OPELM of improvement, Optimally-Pruned Extreme Learning Machine) training study speed and extensive, that robust performance is good advantage, propose a kind of IGBT method for predicting residual useful life based on VKOPP algorithm.Obtain the characterization parameter of collection emitter-base bandgap grading saturation voltage drop as IGBT module failure procedure by accelerated life test, predict by the data-driven Forecasting Methodology that the present invention proposes it, result shows, this method has good estimated performance.

Accompanying drawing explanation

Fig. 1 is a kind of embodiment process flow diagram of IGBT method for predicting residual useful life of the present invention;

Fig. 2 is the experiment raw data of 6 groups of IGBT collection emitter-base bandgap grading saturation voltage drops that accelerated life test obtains;

Fig. 3 is that a kind of embodiment of IGBT method for predicting residual useful life of the present invention predicts the outcome figure, wherein real data (dotted portion), predicted data (bold portion), APT represents acceptable performance threshold, namely failure threshold (wherein, concentrate often organizing IGBT experimental data, first three point two data be used as training set, the data of rear 1/3rd are used as test set).

Embodiment

Below in conjunction with accompanying drawing, the specific embodiment of the present invention is described, so that those skilled in the art understands the present invention better.Requiring particular attention is that, in the following description, when perhaps the detailed description of known function and design can desalinate main contents of the present invention, these are described in and will be left in the basket here.

The present invention need not carry out model construction to concrete circuit component, but based on circuit electric network theory, utilizes the parameter measurability of electric network, adopts theory calculate and actual measurement to combine and designs.

Fig. 1 is a kind of embodiment process flow diagram of IGBT method for predicting residual useful life of the present invention.

In the present embodiment, as shown in Figure 1, IGBT method for predicting residual useful life of the present invention can be divided into four-stage: the acquisition of input data and processing stage, the initial phase of forecast model, saturation voltage drop data prediction stage and IGBT predicting residual useful life stage.Respectively each stage is described in detail below:

(1), input data acquisition and processing stage

S101: by accelerated life test, obtains IGBT collection emitter-base bandgap grading saturation voltage drop aging data time series, obtains data set DS={x (1), x (2) ..., x (D) }, wherein D is the data amount check obtained;

S102: the data set DS that step S101 is obtained, first, determine Embedded dimensions d and the delay time T of its best according to the method for differential entropy rate; Secondly, the function windowize in Matlab is utilized to carry out framing reconstruct to data set DS, and each frame data are mapped to d dimensional feature space respectively, block with the mode p of Volterra progression (p>=1) rank, for the terseness of algorithm, the present invention adopts p=2, thus obtains n (n=D-d τ) individual Frame { (X _t,y _t) (t=1,2 ..., n) to export as the input data of training pattern and target, wherein:

X _t＝{1，x(t)，x(t+τ)，…，x(t+(d-1)τ)，

x ²(t)，x(t)x(t+τ)，…x(t)x(t+(d-1)τ)，

x ²(t+τ)，x(t+τ)x(t+2τ)，…，x(t+τ)x(t+(d-1)τ)，

x ²(t+2τ)，x(t+2τ)x(t+3τ)，…，x(t+2τ)x(t+(d-1)τ)

，…，

x ²(t+(d-2)τ)，x(t+(d-2)x(t+(d-1)τ)

x ²(t+(d-1)τ)}；

Y _t＝x(t+dτ)

Its element number is (d+1) (d+2)/2, and target exports data Y _t=x (t+d τ) (t=1,2 ..., n);

S103: utilize every frame input data X that FB algorithm or LARS algorithm obtain step S102 _tdo optimal selection, select and export Y with target _tthe variable that correlativity is higher, the input data after note selection are:

X _t′＝{1,x(t+c ₁τ),x(t+c ₂τ),…,x(t+c _bτ),…,x(t+c _eτ)x(t+c _fτ),…}，

Wherein: c _b, c _e, c _f∈ 0,1 ..., d-1}, and c _e≤ c _f.X _t' dimension size (element number) be m, then might as well remember X _t'=[x _{t, 1}, x _{t, 2}..., x _t,m] ^t(t=1,2 ..., n).

Two, the initial phase of forecast model

S104: extreme learning machine (ELM) model of initialization one containing N (N < D) individual hidden layer node.The data of input layer are the X that step S103 obtains _t' (t=1,2 ..., n).

At the input weights matrix of coefficients of interval [-1,1] interior random initializtion ELM with threshold coefficient matrix $\stackrel{\&OverBar;}{\mathrm{\&theta;}}={\stackrel{\&OverBar;}{\mathrm{\&theta;}}}_s(s=1,2,...,N;i=1,2,...,m).$ According to initialized with calculate hidden layer output matrix $\stackrel{\&OverBar;}{H}=\&lsqb;{\stackrel{\&OverBar;}{h}}_1,{\stackrel{\&OverBar;}{h}}_2,...,{\stackrel{\&OverBar;}{h}}_t,...,{\stackrel{\&OverBar;}{h}}_n\&rsqb;,$ Wherein ${\stackrel{\&OverBar;}{h}}_t={\left[\begin{array}{ccc}g({\mathrm{\&Sigma;}}_{i=1}^m{\stackrel{\&OverBar;}{w}}_{1,i}{x}_{t,i}-{\stackrel{\&OverBar;}{\mathrm{\&theta;}}}_1)& ...& g({\mathrm{\&Sigma;}}_{i=1}^m{\stackrel{\&OverBar;}{w}}_{N,i}{x}_{t,i}-{\stackrel{\&OverBar;}{\mathrm{\&theta;}}}_N)\end{array}\right]}^T;$

S105: by the hidden layer output matrix obtained in step S104 and the target of its correspondence exports Y in step S102 _t(t=1,2 ..., n), utilize MRSR algorithm vectorial element and hidden layer node rearrange, note sequence after hidden layer node be: wherein, subscript 1≤j _i≤ N is the hidden node sequence number before sequence, and subscript 1≤i≤N is the hidden node precedence sequence number after sequence; According to the order sequenced, l hidden layer node, wherein l < N before extracting.

(3) the saturation voltage drop data prediction stage

S106: w=(w is updated to respectively to the hidden layer node input weights and threshold parameter matrix that step S105 retains _s,i) _{l × m}with θ=θ _s(s=1,2 ..., l; I=1,2 ..., m), and the hidden layer output matrix new to the network calculations after beta pruning $H=\&lsqb;h_1,h_2,...,h_j,...,h_n\&rsqb;,$ Wherein $h_j={\left[\begin{array}{ccc}g({\mathrm{\&Sigma;}}_{i=1}^m{w}_{1,i}{x}_{j,i}-{\mathrm{\&theta;}}_1)& ...& g({\mathrm{\&Sigma;}}_{i=1}^m{w}_{l,i}{x}_{j,i}-{\mathrm{\&theta;}}_l)\end{array}\right]}^T;$

S107: adopt the network after solving beta pruning export weights and predict the outcome, its process is as follows:

1). predicting D+1 data, when namely predicting x (D+1), is X according to the initial input of the known forecast model of step S102 and S103 _t, now, t=D+1-d τ; When after input selection, note is input as X _t' (t=D+1-d τ), then can calculate the network hidden layer output vector in this moment $h_t={\left[\begin{array}{ccc}g({\&Sigma;}_{i=1}^m{w}_{1,i}{x}_{t,i}-{\mathrm{\&theta;}}_1)& ...& g({\&Sigma;}_{i=1}^m{w}_{l,i}{x}_{t,i}-{\mathrm{\&theta;}}_l)\end{array}\right]}^T;$

2). calculate h _tthe matrix H obtained with step S106=[h ₁, h ₂..., h _j..., h _n] in each vectorial h _j(j=1,2 ..., Euclidean distance S n);

3). to h in S _twith each h of H _j(j=1,2 ..., n) distance value sorts, and finds out closest to h from H _tthe new matrix of the individual vector composition of z (z>=l) and the target of correspondence exports ${\hat{Y}}_{z\&times;1}={\&lsqb;{\hat{Y}}_1,{\hat{Y}}_2,...,{\hat{Y}}_z\&rsqb;}^T,$ Thus calculate and export weights accordingly $\mathrm{\&gamma;}={\&lsqb;{\mathrm{\&gamma;}}_1,{\mathrm{\&gamma;}}_2,...,{\mathrm{\&gamma;}}_l\&rsqb;}^T={{\hat{H}}_{z\&times;l}}^{+}{\hat{Y}}_{z\&times;1}={\left({\hat{H}}^T\hat{H}\right)}^{-1}{\hat{H}}^T\hat{Y}.$

The t=D+1-d τ moment is obtained, the one-step prediction value of forecast model by these output weights:

Four, the IGBT predicting residual useful life stage

S108: whether the predicted value that determining step S107 obtains reaches the performance failure threshold value A PT of IGBT module, if do not reached, first, the method using metabolism to fill vacancies in the proper order upgrades the input data frame of IGBT forecast model, the afterbody of initial data frame is added to by new predicted value, remove the header data value of initial data frame simultaneously, ensure that new data frame length is consistent with initial data frame like this.Then 1 in step S107 is repeated), 2), 3) and, constantly update the network hidden layer output vector h in corresponding moment _twith its output weights γ.

If the predicted value of D+W (W>=1) individual data equal or exceed its failure threshold APT, then record this critical point D+W.So prediction obtains W cycle count period IGBT pipe in addition will be in failure state.

In order to technique effect of the present invention is described, altogether completes 6 groups of IGBT accelerated life tests and verify.Fig. 2 tests 6 groups of IGBT collection emitter-base bandgap grading saturation voltage drop aging data time serieses obtaining as experiment raw data EX_1 ~ 6.

Observe the raw data shown in Fig. 2, can find that its data not only exist bad point, also there is the feature of redundancy and higher-dimension, if directly applied to life prediction, then greatly can reduce forecasting efficiency.Therefore, according to the actual needs with the characteristic of problem, in the present embodiment, then compressed transform is carried out to extracted raw data, in lower dimensional space, obtain optimal characteristics, be beneficial to the life prediction realizing IGBT more accurately.Fig. 3 (dotted portion) then depicts the data and curves after process.

In the optimum configurations of forecast model, the method for differential entropy rate is adopted to choose optimum Embedded dimensions and time delay; The nearest-neighbors number arranging model is the exponent number of 100, Volterra progression is three kinds of neuron excitation functions that 2, OPELM uses mixing, and the number of hidden layer neuron is set to 100.

Under above-mentioned parameter condition, for verifying the estimated performance of method proposed by the invention better, different pieces of information number is then utilized to carry out forecast analysis to the residual life of IGBT below, often organize IGBT experimental data concentrate, first three point two data be used as training set, the data of rear 1/3rd are used as test set.(note: training data here and hereafter counts/and training set refers to that the raw data that the input of tectonic model in step S102 and target export, test set then refer to step S107 and data to be predicted in step afterwards.)

Table 1 is the present invention's predicting the outcome under experiment (1).

Table 1

When predicting the residual life of IGBT with the IGBT method for predicting residual useful life that the present invention proposes as can be seen from Table 1, its relative error is all no more than 1%.And also can obtain when using a small amount of data in forecast model and predict the outcome preferably.

As can be seen from Figure 3, IGBT method for predicting residual useful life of the present invention not only can well predict the overall trend of IGBT experimental data, can also more adequately dope its remaining lifetime value.

Although be described the illustrative embodiment of the present invention above; so that those skilled in the art understand the present invention; but should be clear; the invention is not restricted to the scope of embodiment; to those skilled in the art; as long as various change to limit and in the spirit and scope of the present invention determined, these changes are apparent, and all innovation and creation utilizing the present invention to conceive are all at the row of protection in appended claim.

Claims (1)

1. an IGBT method for predicting residual useful life, is characterized in that, comprises the following steps:

(1), by accelerated life test, obtain IGBT collection emitter-base bandgap grading saturation voltage drop aging data time series, obtain data set DS={x (1), x (2) ..., x (D) }, wherein D is the data amount check obtained;

(2), first, for data set DS, best Embedded dimensions d and delay time T is determined according to the method for differential entropy rate; Then, framing reconstruct is carried out to data set DS, and each frame data are mapped to d dimensional feature space respectively, thus obtain n (n=D-d τ) individual Frame { (X _t,y _t), wherein t=1,2 ..., n, to export as the input data of training pattern and target, wherein:

X _t＝{1,x(t),x(t+τ),…,x(t+(d-1)τ),

x ²(t),x(t)x(t+τ),…x(t)x(t+(d-1)τ),

x ²(t+τ),x(t+τ)x(t+2τ),…,x(t+τ)x(t+(d-1)τ),

x ²(t+2τ),x(t+2τ)x(t+3τ),…,x(t+2τ)x(t+(d-1)τ)

…，

x ²(t+(d-2)τ),x(t+(d-2)x(t+(d-1)τ)

x ²(t+(d-1)τ)}；

Y _t＝x(t+dτ)

(3) Forward-backward algorithm (FB, is utilized, Forward-Backward algorithm) or minimum angle regression algorithm (LARS, Least Angle Regression algorithm) every frame input data X that step (2) is obtained _tdo optimal selection, select and export Y with corresponding target _tm the input data that correlativity is higher, the every frame input data after note selection are:

X ' _t={ 1, x (t+c ₁τ), x (t+c ₂τ) ..., x (t+c _bτ) ..., x (t+c _eτ) x (t+c _fτ) }; c _b, c _e, c _f∈ 0,1 ..., d-1}, and c _e≤ c _f;

Then every frame input data X ' _tdimension size (element number) be m, can again be expressed as:

X′ _t＝[x _t,1,x _t,2,…,x _t,m] ^T(t＝1,2,…,n)；

(4), extreme learning machine (ELM) model of initialization one containing N (N < D) individual hidden layer node; The data of input layer are the input data X ' that step (3) obtains _t(t=1,2 ..., n);

At the input weights matrix of coefficients of interval [-1,1] interior random initializtion ELM with threshold coefficient matrix wherein s=1,2 ..., N; I=1,2 ..., m;

According to initialized input weights matrix of coefficients with threshold coefficient matrix calculate hidden layer output matrix wherein network hidden layer output vector for:

${\stackrel{\&OverBar;}{h}}_t={\&lsqb;g({\&Sigma;}_{i=1}^m{\stackrel{\&OverBar;}{w}}_{1,i}{x}_{t,i}-{\stackrel{\&OverBar;}{\mathrm{\&theta;}}}_1),g({\&Sigma;}_{i=1}^m{\stackrel{\&OverBar;}{w}}_{2,i}{x}_{t,i}-{\stackrel{\&OverBar;}{\mathrm{\&theta;}}}_2),\mathrm{...},g({\&Sigma;}_{i=1}^m{\stackrel{\&OverBar;}{w}}_{N,i}{x}_{t,i}-{\stackrel{\&OverBar;}{\mathrm{\&theta;}}}_N)\&rsqb;}^T,$ G represents excitation function;

(5), according to the vector that step (4) obtains and the corresponding target obtained in step (2) exports Y _t, utilize multiple response sparse regression algorithm (MRSR, Multiresponse Sparse Regression) to vector element and hidden layer node rearrange, note sequence after hidden layer node be:

wherein, subscript 1≤j _i≤ N is the hidden layer node sequence number before sequence, and subscript 1≤i≤N is the hidden layer node precedence sequence number after sequence; According to the order sequenced, l hidden layer node, wherein l < N before extracting;

(6), according to step (5) extract input weights coefficient corresponding to hidden layer node and threshold coefficient, input weights coefficient and threshold coefficient matrix are updated to w=(w respectively _s,i) _{l × m}with θ=θ _ss=1,2 ..., l; I=1,2 ..., m, and to after extraction (beta pruning), recalculate hidden layer output matrix H=[h ₁, h ₂..., h _j..., h _n], wherein vector $h_j={\left[\begin{array}{ccc}g({\&Sigma;}_{i=1}^m{w}_{1,i}{x}_{j,i}-{\mathrm{\&theta;}}_1)& \mathrm{...}& g({\&Sigma;}_{i=1}^m{w}_{l,i}{x}_{j,i}-{\mathrm{\&theta;}}_l)\end{array}\right]}^T;$

(7), solve beta pruning after network export weights and predicting the outcome, its process is as follows:

7.1), predict D+1 data, namely time predicted data x (D+1), the original input data frame according to step (2) and (3) known forecast model is X _t, now, t=D+1-d τ; When after input selection, note selects rear input data frame to be X ' _t, t=D+1-d τ, then can calculate the network hidden layer output vector in this moment and t=D+1-d τ moment $h_t={\left[\begin{array}{ccc}g({\&Sigma;}_{i=1}^m{w}_{1,i}{x}_{t,i}-{\mathrm{\&theta;}}_1)& \mathrm{...}& g({\&Sigma;}_{i=1}^m{w}_{l,i}{x}_{t,i}-{\mathrm{\&theta;}}_l)\end{array}\right]}^T;$

7.2), computational grid hidden layer output vector h _tthe hidden layer output matrix H=[h obtained with step (6) ₁, h ₂..., h _j..., h _n] in each vectorial h _j, j=1,2 ..., the Euclidean distance S of n;

7.3), to network hidden layer output vector h in Euclidean distance S _twith each vectorial h of hidden layer output matrix H _j, j=1,2 ..., the distance value of n sorts, and finds out closest to network hidden layer output vector h from hidden layer output matrix H _tz, z>=l the new matrix of vector composition and the target of correspondence exports ${\hat{Y}}_{z\&times;1}={\&lsqb;{\hat{Y}}_1,{\hat{Y}}_2,\mathrm{...},{\hat{Y}}_z\&rsqb;}^T,$ Thus calculating exports weights accordingly $\mathrm{\&gamma;}={\&lsqb;{\mathrm{\&gamma;}}_1,{\mathrm{\&gamma;}}_2,\mathrm{...},{\mathrm{\&gamma;}}_l\&rsqb;}^T={{\hat{H}}_{z\&times;l}}^{+}{\hat{Y}}_{z\&times;1}={\left({\hat{H}}^T\hat{H}\right)}^{-1}{\hat{H}}^T\hat{Y};$

The t=D+1-d τ moment is obtained, the predicted value of forecast model by these output weights:

(8), the predicted value that obtains of determining step (7) whether reach the performance failure threshold value (APT, Acceptable Performance Threshold) of IGBT, if reached, then this predicted value position be D+1 as critical point record, stop prediction, if do not reached, then by predicted value join IGBT collection emitter-base bandgap grading saturation voltage drop aging data time series, obtain data set DS={x (2), x (3) ..., x (D+1) }, then obtain predicted value according to the method for step (1) ~ (7) and judge whether the performance failure threshold value reaching IGBT, if no, then continue to join IGBT collection emitter-base bandgap grading saturation voltage drop aging data time series, obtain data set DS={x (3), x (4),, x (D+2) }, repeat like this, until reach, and using this predicted value position D+W as critical point record, wherein, W is predicted value reach the prediction number of times of the performance failure threshold value of IGBT.