IGBT residual life prediction method based on variational Bayes and hybrid Wiener process

Abstract

The invention discloses an IGBT residual life prediction method based on variational Bayes and a mixed Wiener process, and belongs to the technical field of IGBT residual life prediction, and the method comprises the following steps: S1, giving out the definition of the mixed Wiener process, and building a degradation model to describe the degradation track of an IGBT; s2, performing off-line parameter estimation based on a variational Bayesian algorithm by using a prior sample; s3, selecting an optimal model in the degradation models according to an automatic correlation determination method; s4, selecting a prior sample which is most similar to the test sample at the current moment; s5, updating the mean value and variance of the weight vector of the test sample on line; and S6, solving a probability density function of the residual life of the mixed Wiener process in the sense of the first arrival time, and estimating the residual life value of the IGBT at each moment. According to the IGBT residual life prediction method based on the variational Bayes and the mixed Wiener process, the degradation mode can be automatically determined, the fitting precision of an IGBT degradation track is improved, the requirement for specialty is reduced, and the universality of the model is improved.

Description

Technical Field

[0001] The invention relates to the technical field of IGBT remaining life prediction, and in particular to an IGBT remaining life prediction method based on variational Bayes and hybrid Wiener process. Background Art

[0002] With the continuous development of modern society, higher safety and reliability requirements are also put forward for infrastructure equipment. As the core component that determines the starting and braking speed, the adjustment of the running speed and the efficiency of energy utilization of high-speed rail, accurate remaining life prediction of insulated gate bipolar transistor (IGBT) chips is a key task, which can improve the reliability of system operation, reduce the risk of failure and reduce economic costs. For IGBT, its degradation process may lead to more complex multi-stage degradation due to changes in internal mechanisms or external working conditions. Existing studies have selected one of several typical models as the degradation mode, which may lead to inaccurate description of the IGBT degradation trajectory. Therefore, it is necessary to study a method that automatically determines the form of the degradation mode without manual specification, and can select one or more drift terms at the same time and merge them into a degradation mode that can describe the complex degradation trajectory. In addition, considering the ease of use and practicality, the proposed model should maintain a high degree of simplicity and interpretability, and the parameters need to be updated online to reflect the differences between individuals and reduce the demand for the number of historical samples. Therefore, there is an urgent need for an IGBT remaining life prediction method that can meet the above requirements. Summary of the invention

[0003] The purpose of the present invention is to provide an IGBT remaining life prediction method based on variational Bayes and hybrid Wiener process, which can automatically determine the degradation mode and describe complex degradation trajectories, improve the fitting accuracy of the IGBT degradation trajectory, reduce the requirements for professionalism, and improve the versatility of the model.

[0004] To achieve the above object, the present invention provides an IGBT remaining life prediction method based on variational Bayesian and hybrid Wiener process, comprising the following steps:

[0005] S1. Give the definition of the hybrid Wiener process and establish a degradation model to describe the degradation trajectory of IGBT;

[0006] S2, using prior samples to perform offline parameter estimation based on the variational Bayes algorithm;

[0007] S3, selecting the optimal model in the degradation model according to the automatic correlation determination method;

[0008] S4, select the prior sample that is most similar to the test sample at the current moment;

[0009] S5, update the mean and variance of the weight vector of the test sample online;

[0010] S6. Calculate the probability density function of the remaining life of the mixed Wiener process in the sense of first arrival time, and estimate the remaining life value of the IGBT at each moment.

[0011] Preferably, S2 specifically includes the following steps:

[0012] S2.1. According to the nth prior sample t k Time and t k-1 The degradation model at time Δx is used to calculate the difference in degradation amount. nk Expression of

[0013] S2.2. Find Δx nk Vector The likelihood function of ;

[0014] S2.3, normalize all data from the same sensor to the same scale so that the maximum value is 1;

[0015] S2.4. Give the distribution of prior parameters and conjugate prior distribution;

[0016] S2.5, updating the variational decomposition of each unknown parameter in turn according to the variational Bayes algorithm;

[0017] S2.6, calculate the variational lower bound ELBO;

[0018] S2.7. Iterate the parameter update until the difference between the variational lower bounds of two iterations is less than the preset threshold or the number of iterations is greater than 1000. Then stop the iteration and obtain the final parameter estimate.

[0019] Preferably, S3 specifically includes the following steps:

[0020] S3.1, based on the current parameter estimates obtained in S2.7, complete model selection according to the automatic correlation determination method;

[0021] S3.2, iterate the parameter estimation of S2 and the model selection of S3.1 until only one model term remains;

[0022] S3.3. Select the term vector corresponding to the largest variational lower bound value as the optimal model.

[0023] Preferably, S4 specifically comprises the following steps:

[0024] S4.1. Calculate the Euclidean distance and DTW distance between the degradation sequence of the test sample and all prior samples from the initial moment to the current moment respectively;

[0025] S4.2, calculate the score of each prior sample at the current moment;

[0026] S4.3. Select the prior sample corresponding to the smallest score as the most similar sample to the test sample at the current moment.

[0027] Preferably, S5 specifically includes the following steps:

[0028] S5.1. Calculate the posterior distribution of the weight vector at the current moment based on the prior parameter distribution of the most similar training sample and the sequential Bayesian algorithm;

[0029] S5.2. Simplify the posterior distribution of the weight vector and update the mean and variance of the weight vector of the test sample online.

[0030] Therefore, the present invention adopts the above-mentioned IGBT remaining life prediction method based on variational Bayes and hybrid Wiener process, which can automatically determine the degradation mode and describe complex degradation trajectories, improve the fitting accuracy of IGBT degradation trajectories, reduce the requirements for professionalism, and improve the versatility of the model.

[0031] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a diagram of an IGBT remaining life prediction method based on variational Bayes and hybrid Wiener process of the present invention;

[0033] Figure 2 It is a degradation path combination diagram of an IGBT remaining life prediction method based on variational Bayes and hybrid Wiener process of the present invention;

[0034] Figure 3 It is a probability density function diagram of the remaining life of an IGBT remaining life prediction method based on variational Bayesian and hybrid Wiener process of the present invention;

[0035] Figure 4 It is a remaining life estimation diagram of an IGBT remaining life prediction method based on variational Bayesian and hybrid Wiener process of the present invention;

[0036] Figure 5 It is a comparison diagram of the absolute error of the remaining life of an IGBT remaining life prediction method based on variational Bayes and hybrid Wiener process of the present invention. DETAILED DESCRIPTION

[0037] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0038] Unless otherwise defined, technical or scientific terms used in the present invention shall have the common meanings understood by one having ordinary skills in the field to which the present invention belongs.

[0039] Embodiment 1

[0040] like Figure 1 As shown, the present invention provides an IGBT remaining life prediction method based on variational Bayes and hybrid Wiener process, comprising the following steps:

[0041] S1. Give the definition of the hybrid Wiener process and establish a degradation model to describe the degradation trajectory of IGBT;

[0042] A degradation model that can automatically determine the form and quantity of drift terms is established, called the mixed Wiener process, and the expression is:

[0043]

[0044] in,

[0045]

[0046] Moreover, X(t) is the degradation value at time t, x0 is the initial degradation value, F(t) is the nonlinear drift term after P functions are weighted, θ is the weight vector, σ B is the diffusion coefficient, B(t) is the standard Brownian motion, and f(·) represents a function of any form. For simplicity, x0 is assumed to be 0. When x0 is non-zero, the assumption can be simply realized by subtracting the initial degradation amount x0 from the degradation value X(t).

[0047] S2. Use prior samples to perform offline parameter estimation based on the variational Bayesian algorithm; specifically, the following steps are included:

[0048] Assume that there are n prior IGBT samples available for training. The full life cycle degradation data of these prior samples is known (here the collector-emitter saturation voltage is selected as the degradation amount), and they are of the same type as the test samples.

[0049] S2.1. According to the nth prior sample t k Time and t k-1 The degradation model at time Δx is used to find the difference in degradation amount nk Expression of

[0050]

[0051] Where ΔX n (t k )=X n (t k )-X n (tk-1 ) and Δt k =t k -t k-1 (n∈{1,2,...,N}), N represents the number of prior samples, k∈{2,3,...,K n}, t k Indicates the current time, K n Represents the total number of state monitoring data of the nth prior sample.

[0052] S2.2. Find Δx nk Vector The likelihood function of ;

[0053] For the convenience of expression, let ΔF nk =F n (t k )-F n (t k-1 ), Δx nk Denotes ΔX n (t k ), Obviously, ΔX n It obeys the multivariate normal distribution, and its likelihood function is:

[0054]

[0055] in, represents a normal distribution about x, with mean and variance ΔF respectively nk and

[0056] S2.3, normalize all data from the same sensor to the same scale so that the maximum value is 1;

[0057] Let Δf nk =f n (t k )-f n (t k-1 ), then ΔF nk =Δf nk θ n ; Then, all data from the same sensor are normalized to the same scale so that the maximum value is 1, that is: Δf p (k) = Δf p (k) / Zoom(p), where Δf p (k) = f p (k)-f p (k-1) and Zoom(p) = max{Δf p (1),Δf p (2),...,Δf p (KN )}; In addition, due to the incomplete degradation data of the test samples in the online stage, the Zoom(p) of the preprocessed corresponding most similar prior sample is selected as its normalization scale.

[0058] S2.4. Give the distribution of prior parameters and conjugate prior distribution;

[0059] variance The conjugate prior distribution of is the inverse Gamma distribution. For the convenience of calculation, ξ n express The reciprocal of ; and the commonly used joint conjugate prior assumptions for normal distribution and Gamma distribution are:

[0060]

[0061] Among them, G(ξ n |α0,β0) represents the Gamma distribution about x, whose shape parameter and scale parameter are α0 and β0 respectively. is the corresponding hyperparameter,

[0062] Then, in order to simplify the computational complexity, the Gamma distribution is introduced as Priors:

[0063]

[0064] Among them, δ0 and γ0 are hyper parameters.

[0065] S2.5, updating the variational decomposition of each unknown parameter in turn according to the variational Bayes algorithm;

[0066] According to the variational Bayes method, the variational solutions of the unknown parameters at the mth iteration are:

[0067] 1)q m (θ n ,ξ n ) has the same form as its joint conjugate prior, which is:

[0068]

[0069] where {θ nm ,V nm ,α nm ,β nm The value of} is obtained by the following calculation:

[0070]

[0071] 2) is the product of a set of independent Gamma distributions:

[0072]

[0073] in δ nm and γ nmp The value of is calculated by the following formula:

[0074]

[0075] {θ nm ,V nm ,α nm ,β nm Unknown term in expression is a diagonal matrix, we have:

[0076]

[0077] In the process of updating unknown parameters, the variables The initial values ​​of

[0078] S2.6, calculate the variational lower bound ELBO;

[0079]

[0080] S2.7. Iterate the parameter update until the difference between the variational lower bounds of two iterations is less than the preset threshold or the number of iterations is greater than 1000. Then stop the iteration and obtain the final parameter estimate.

[0081] S3, selecting the optimal model in the degradation model according to the automatic correlation determination method; specifically comprising the following steps:

[0082] S3.1, based on the current parameter estimates obtained in S2.7, complete model selection according to the automatic correlation determination method;

[0083] Based on the automatic correlation determination method, the minimum The corresponding terms are regarded as irrelevant terms, that is, θ nm ≈0 is not considered as an item in the optimal degradation model. for:

[0084] exp{min(lnS n )+[max(lnS n )-min(lnS n )] / b};

[0085] Among them, S n ={S n1 ,S n2 ,...,S nP}、S nprepresents the auto-correlation determined value, b is a tuning parameter, and b=10 is selected.

[0086] S3.2, iterate the parameter estimation of S2 and the model selection of S3.1 until only one model term remains;

[0087] S3.3. Select the term vector corresponding to the largest variational lower bound value as the optimal model.

[0088] S4, selecting the prior sample that is most similar to the test sample at the current moment; specifically comprising the following steps:

[0089] S4.1. Calculate the Euclidean distance and DTW distance between the degradation sequence of the test sample and all prior samples from the initial moment to the current moment respectively;

[0090] 1) Euclidean distance ED nk :

[0091]

[0092] where x nk Indicates the nth prior sample at time t k The degradation amount, k now =min(k,K n ).

[0093] 2) DTW distance DTW nk :

[0094] DTW distance uses dynamic programming to align two degradation sequence value curves and calculates the similarity between two curves of different lengths. DTW distance pays more attention to the trend and shape of the time series, which cannot be solved by Euclidean distance. and the degradation value sequence of the nth prior sample x a and x nb The DTW distance between them is:

[0095] DTW(x a ,x nb )=(x a -x nb ) 2 ;

[0096] in,

[0097] The curved path can be obtained by using the distance matrix M DTW The elements in the matrix are M DTW The shortest path in is:

[0098] D a,b =DTW(x a ,x nb )+min{D a-1,b-1 ,D a-1,b .D a,b-1};

[0099] Among them, D a,b represents the cumulative distance, in particular, D 1,1 =DTW(x1,x n1 ) holds true. In this way, the similarity based on DTW distance can be obtained by calculate.

[0100] S4.2, calculate the score of each prior sample at the current moment;

[0101]

[0102] S4.3. Select the prior sample corresponding to the smallest score as the most similar sample to the test sample at the current moment.

[0103] S5. Online updating of the mean and variance of the weight vector of the test sample; specifically comprising the following steps:

[0104] S5.1. Calculate the posterior distribution of the weight vector at the current moment based on the prior parameter distribution of the most similar training sample and the sequential Bayesian algorithm;

[0105]

[0106] Where, represents the time t k The degenerate state x k (k>1).

[0107] S5.2, simplify the posterior distribution of the weight vector, and update the mean and variance of the weight vector of the test sample online;

[0108] θ follows a normal distribution with a mean of Ω k and variance H k It can be updated by:

[0109]

[0110] in, Ω k and H k The initial values ​​are Ω1=θ nm and

[0111] S6. Calculate the probability density function of the remaining life of the mixed Wiener process in the sense of first arrival time, and estimate the remaining life value of the IGBT at each moment.

[0112] Based on the proposed hybrid Wiener process degradation model, the probability density function expression of the remaining life of the IGBT is:

[0113]

[0114] Where ω represents the given failure threshold, L k express The corresponding remaining life, F′(t) is the first-order derivative of F(t) with respect to t.

[0115] Based on the above formula, the probability density function of the remaining life of the IGBT at each moment can be calculated, and its maximum value is the estimated value of the remaining life.

[0116] In order to verify the performance of the IGBT remaining life prediction method, NASA's IGBT accelerated aging dataset is selected for illustration, and the prediction results are compared with other advanced methods. Figure 2 a in the figure gives the degradation trajectory of the IGBT sample in the dataset. The degradation amount is selected as the collector-emitter saturation voltage, and the failure threshold is set to be 5.5V higher than the initial value. The test sample uses DEVICE4, and the training samples are DEVICE2 / DEVICE3. Figure 2 b in the figure is the degradation trajectory of the test sample and the estimated degradation trajectory fitted by the hybrid Wiener degradation model and parameter estimation method proposed in the present invention, which shows that the degradation model of the present invention can better fit the real degradation trajectory. Figure 3 The probability density function of the remaining life is given by Figure 4 The average estimated value of the remaining life of the IGBT is given. Figure 5 A comparison chart of the absolute error between the true value and the estimated value of the remaining life of this method and the multi-stage nonlinear Wiener process method is given. In addition, the root mean square error and mean absolute error are used to quantify the prediction performance of this method. The root mean square errors of this method and the multi-stage nonlinear Wiener process method are 14.38 and 19.75, respectively, and the mean absolute errors are 8.67 and 17.40, respectively. It can be seen that the method of the present invention shows a more superior performance.

[0117] Therefore, the present invention adopts the above-mentioned IGBT remaining life prediction method based on variational Bayes and hybrid Wiener process, which can automatically determine the degradation mode and describe complex degradation trajectories, improve the fitting accuracy of IGBT degradation trajectories, reduce the requirements for professionalism, and improve the versatility of the model.

[0118] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A method for predicting the remaining life of an IGBT based on variational Bayesian and hybrid Wiener process, characterized in that: The following steps are involved: S1. Give the definition of the hybrid Wiener process and establish a degradation model to describe the degradation trajectory of IGBT; S2, using prior samples to perform offline parameter estimation based on the variational Bayes algorithm; S3, selecting the optimal model in the degradation model according to the automatic correlation determination method; S4, select the prior sample that is most similar to the test sample at the current moment; S5, update the mean and variance of the weight vector of the test sample online; S6. Calculate the probability density function of the remaining life of the mixed Wiener process in the sense of first arrival time, and estimate the remaining life value of the IGBT at each moment.

2. The IGBT remaining life prediction method based on variational Bayesian and hybrid Wiener process according to claim 1, characterized in that: S2 specifically includes the following steps: S2.

1. According to the nth prior sample t k Time and t k-1 The degradation model at time Δx is used to calculate the difference in degradation amount. nk Expression of S2.

2. Find Δx nk The vector composed of The likelihood function of ; S2.3, normalize all data from the same sensor to the same scale so that the maximum value is 1; S2.

4. Give the distribution of prior parameters and conjugate prior distribution; S2.5, updating the variational decomposition of each unknown parameter in turn according to the variational Bayes algorithm; S2.6, calculate the variational lower bound ELBO; S2.

7. Iterate the parameter update until the difference between the variational lower bounds of two iterations is less than the preset threshold or the number of iterations is greater than 1000. Then stop the iteration and obtain the final parameter estimate.

3. The IGBT remaining life prediction method based on variational Bayesian and hybrid Wiener process according to claim 2, characterized in that: S3 specifically includes the following steps: S3.1, based on the current parameter estimates obtained in S2.7, complete model selection according to the automatic correlation determination method; S3.2, iterate the parameter estimation of S2 and the model selection of S3.1 until only one model term remains; S3.

3. Select the term vector corresponding to the largest variational lower bound value as the optimal model.

4. The IGBT remaining life prediction method based on variational Bayesian and hybrid Wiener process according to claim 1, characterized in that: S4 specifically includes the following steps: S4.

1. Calculate the Euclidean distance and DTW distance between the degradation sequence of the test sample and all prior samples from the initial moment to the current moment respectively; S4.2, calculate the score of each prior sample at the current moment; S4.

3. Select the prior sample corresponding to the smallest score as the most similar sample to the test sample at the current moment.

5. The IGBT remaining life prediction method based on variational Bayesian and hybrid Wiener process according to claim 1, characterized in that: S5 specifically includes the following steps: S5.

1. Calculate the posterior distribution of the weight vector at the current moment based on the prior parameter distribution of the most similar training sample and the sequential Bayesian algorithm; S5.

2. Simplify the posterior distribution of the weight vector and update the mean and variance of the weight vector of the test sample online.

Images