CN114707423A - Method for predicting residual life of IGBT (insulated Gate Bipolar transistor) - Google Patents

Method for predicting residual life of IGBT (insulated Gate Bipolar transistor) Download PDF

Info

Publication number
CN114707423A
CN114707423A CN202210456063.2A CN202210456063A CN114707423A CN 114707423 A CN114707423 A CN 114707423A CN 202210456063 A CN202210456063 A CN 202210456063A CN 114707423 A CN114707423 A CN 114707423A
Authority
CN
China
Prior art keywords
igbt
aging
data
residual life
elm
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
Application number
CN202210456063.2A
Other languages
Chinese (zh)
Inventor
刘伯颖
李忠涛
陈国龙
李玲玲
刘佳琪
周超
王海宇
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Hebei University of Technology
Original Assignee
Hebei University of Technology
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Hebei University of Technology filed Critical Hebei University of Technology
Priority to CN202210456063.2A priority Critical patent/CN114707423A/en
Publication of CN114707423A publication Critical patent/CN114707423A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
    • G06F2119/04—Ageing analysis or optimisation against ageing
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02E—REDUCTION OF GREENHOUSE GAS [GHG] EMISSIONS, RELATED TO ENERGY GENERATION, TRANSMISSION OR DISTRIBUTION
    • Y02E10/00—Energy generation through renewable energy sources
    • Y02E10/70—Wind energy
    • Y02E10/76—Power conversion electric or electronic aspects

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Medical Informatics (AREA)
  • Software Systems (AREA)
  • Artificial Intelligence (AREA)
  • Computer Hardware Design (AREA)
  • Geometry (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention relates to a method for predicting the residual life of an IGBT (insulated gate bipolar translator), in particular to a method for predicting the residual life of an IGBT based on an improved runoff flow direction algorithm optimization extreme learning machine, which is technically characterized in that: obtaining IGBT aging test data including junction temperature data, aging cycle frequency data and saturation pressure drop data through an IGBT power cycle aging test, preprocessing the aging data of the IGBT and analyzing the correlation of the aging data and the health state of the IGBT, establishing an IGBT residual life prediction model based on an improved runoff flow direction algorithm optimization extreme learning machine, and finally substituting the aging test data of the IGBT into verification of the effectiveness of the IGBT residual life prediction model based on the improved runoff flow direction algorithm optimization extreme learning machine. The method is reasonable in design, and the improved radial flow direction algorithm optimized extreme learning machine (IFDA-ELM) model is found to have higher effectiveness and stability in the aspect of predicting the residual life of the IGBT through a test result. The method can be used for better health state evaluation and residual life prediction of the IGBT, and is a new exploration of an information fusion method in the aspect of life prediction.

Description

Method for predicting residual life of IGBT
Technical Field
The technical scheme of the invention belongs to the technical field of reliability of power electronic devices (IGBT), and relates to a method for predicting the residual life of the IGBT.
Background
In the 21 st century, excessive consumption of traditional energy sources such as petroleum, natural gas and coal has caused environmental pollution and global temperature rise. Therefore, various countries have turned development points to renewable, low-pollution, and widely distributed new energy sources, such as wind energy, solar energy, tidal energy, and the like. Thus, new energy power generation technology has been developed. The new energy power generation can not be separated from the current transformation conversion function of the power electronic current transformation device, and the reliability of the new energy power generation plays an extremely important role in the overall stable operation of the system. In wind power generation, due to uncertainty of external factors, the converter device is difficult to be in a normal operation state, and aging failure of the converter device is accelerated. If the converter device has problems, the reliability of the operation of the system is influenced, and even the system is broken down. Therefore, the reliability of the inverter device must be considered.
At present, IGBTs have been used in many fields and play an important role therein, and the reliable operation of an IGBT affects the operational reliability of the system in which the IGBT is located. In a power electronic converter system, about 34% of faults are caused by power semiconductor device failures; the failure of a photovoltaic inverter with an IGBT as core accounts for about 37% of the total failure of the system. Once the IGBT fails, the power electronic converter system may malfunction, even cause system breakdown, and cause economic loss and casualties. Thus, the reliability of IGBTs limits the development of power and other systems. The reliable operation of the IGBT is related to the quality of life of people, the stability of society and the development of national economy, so it is necessary to explore the reliability research method of the IGBT. The reliability research method of the IGBT mainly comprises IGBT state monitoring, IGBT health state evaluation, residual life prediction and the like. In the method for researching the reliability of the IGBT, the aging parameters of the IGBT are generally introduced into the method for researching the reliability of the IGBT. The health state of the IGBT is estimated by using the aging parameters and combining a mathematical model. According to the IGBT health state evaluation result, the operation conditions of the IGBT in various states can be mastered, and whether the operation of the IGBT is normal or not is judged. And predicting the residual life of the IGBT by the time sequence of the IGBT aging parameters. According to the prediction result of the residual service life of the IGBT, the IGBT can be replaced before failure, so that the maintenance cost of the system with the IGBT is reduced, the probability of failure occurrence is reduced, and the reliability of the system is improved.
The existing RUL prediction capability still has some problems, such as high complexity of a magneto-electric-thermal analysis model, high simulation precision of the model and high time cost; the health state of the IGBT can be fully researched by applying a probability statistical method, the probability distribution of IGBT failure can be fully researched, certain advantages are achieved in the aspect of evaluation efficiency, time cost is saved, but the evaluation precision is not satisfactory, and therefore the traditional RUL prediction capability still needs to be improved.
Disclosure of Invention
The invention aims to overcome the defects of the prior art, provides a method for predicting the residual life of an IGBT, and particularly relates to a method for predicting the residual life of an IGBT based on an improved runoff flow direction algorithm optimization extreme learning machine, so as to predict the residual life of the IGBT; obtaining IGBT aging test data including junction temperature data, aging cycle frequency data and saturation pressure drop data through an IGBT power cycle aging test, preprocessing the aging data of the IGBT and analyzing the correlation of the aging data and the health state of the IGBT, establishing an IGBT residual life prediction model based on an improved runoff flow direction algorithm optimization extreme learning machine, and finally substituting the aging test data of the IGBT into verification of the effectiveness of the IGBT residual life prediction model based on the improved runoff flow direction algorithm optimization extreme learning machine. The invention solves the technical problems in the prior art by adopting the following technical scheme:
a method for predicting the residual life of an IGBT (insulated gate bipolar transistor), in particular to a method for predicting the residual life of an IGBT based on an improved runoff flow direction algorithm optimization extreme learning machine, which comprises the following steps:
step 1, obtaining IGBT aging test data through an IGBT power cycle aging test, wherein the IGBT aging test data comprises junction temperature data, aging cycle number data and saturation voltage drop data;
step 2, carrying out aging data pretreatment on the IGBT and analyzing the correlation of the IGBT with the health state of the IGBT;
step 3, establishing an IGBT residual life prediction model based on an improved runoff flow direction algorithm optimization extreme learning machine (IFDA-ELM);
and 4, substituting the aging test data of the IGBT into verification of the effectiveness of the IGBT residual life prediction model based on the improved runoff flow direction algorithm optimization extreme learning machine.
2. The IGBT residual life prediction method based on the improved runoff flow direction algorithm optimization extreme learning machine as claimed in claim 1, characterized in that: the specific implementation method of the step 1 comprises the following steps: in the prediction of the residual life of the IGBT, under a certain aging condition, firstly, the aging of a module is accelerated; secondly, after the module is aged to a certain degree, the module is subjected to static discrete data measurement; and finally, acquiring parameter data through a single-pulse measurement test, and acquiring steady-state thermal resistance data, junction temperature data and aging cycle number data of the module through a thermal resistance measurement test.
3. The IGBT residual life prediction method based on the improved runoff flow direction algorithm optimization extreme learning machine as claimed in claim 1, characterized in that: the specific implementation method of the step 2 comprises the following steps:
step 2.1, preprocessing the data of IGBT residual life prediction by using wavelet denoising;
and 2.2, aiming at the correlation problem of the aging parameters and the health state of the IGBT, the aging parameter data of the IGBT is obtained through an accelerated aging test, and the mapping relation between the aging parameter set of the IGBT and the health state grade of the IGBT is established based on the aging parameter data analysis.
4. The IGBT residual life prediction method based on the improved runoff flow direction algorithm optimization extreme learning machine as claimed in claim 1, characterized in that: the specific implementation method of the step 3 comprises the following steps:
and 3.1, evaluating the health state of the IGBT based on the extreme learning machine. The ELM is an algorithm proposed on the basis of a single hidden layer feedforward neural network, and an Extreme Learning Machine (ELM) can determine the mapping relation between the input quantity and the output quantity of the IGBT by learning the IGBT aging test data. Then, the aging parameter set of the IGBT is used as the input of the ELM, and the number of IGBT aging cycles is used as the output of the ELM. Obtaining the relation between the aging parameter set and the aging cycle times through model training; and obtaining a predicted value of the aging cycle times through testing, and accurately judging the health state grade of the IGBT according to the IGBT health state evaluation standard.
And 3.2, because the evaluation effect of the Extreme Learning Machine (ELM) is influenced by the initial weight and the threshold value, but the improved runoff flow direction algorithm (IFDA) has the advantages of good optimizing effect, strong convergence performance and the like, the invention adopts the IFDA to search the optimal weight and the threshold value of the ELM and establishes an IFDA-ELM-based IGBT health state evaluation model.
5. The IGBT residual life prediction method based on the improved runoff flow direction algorithm optimization extreme learning machine as claimed in claim 1, characterized in that: the specific implementation method of the step 4 comprises the following steps:
step 4.1, importing the test data into an IFDA-ELM model, testing the IFDA-ELM model, and predicting the residual life of the IGBT;
and 4.2, analyzing the prediction result of the residual life of the IGBT based on the IFDA-ELM model.
According to the method, aging failure analysis and aging parameter extraction are carried out on the IGBT through an aging test, then correlation analysis of aging data preprocessing and the health state of the IGBT is carried out, then an IGBT residual life prediction model based on an improved runoff flow direction algorithm optimization extreme learning machine is established, finally, the aging parameters of the IGBT are brought into verification of the effectiveness of the IGBT residual life prediction model based on the improved runoff flow direction algorithm optimization extreme learning machine, the most effective prediction model is obtained, and the method used in the method has a good reference function for similar time sequence data prediction.
The invention has the advantages and positive effects that:
1. the invention provides an improved runoff flow direction algorithm (IFDA) based on aging information of an IGBT. Compared with other traditional algorithms, the algorithm has the advantages of higher optimization precision and higher convergence speed.
2. The method is reasonable in design, and through analysis of test results, the effectiveness and the stability of the IFDA-ELM model on the prediction of the residual life of the IGBT are superior to those of the other three models, so that the effectiveness of the IFDA-ELM model on the prediction of the residual life of the IGBT is verified. Therefore, the IFDA-ELM model is not only suitable for IGBT health state evaluation, but also can be used for IGBT residual life prediction. And in the case of obvious module degradation characteristics, the method has the highest accuracy and is a new exploration of the information fusion method in the aspect of life prediction.
Drawings
FIG. 1 is a flowchart of a method for IGBT lifetime prediction in an embodiment of the present invention;
FIG. 2 is a schematic cross-sectional view of an IGBT module according to an embodiment of the invention;
FIG. 3 is a diagram of the ln (R) function in an embodiment of the present invention;
FIG. 4 is a model of the relationship between aging cycle number, crusting resistance, and health level for an embodiment of the present invention;
FIG. 5 is a diagram illustrating the relationship between the crusting thermal resistance, the number of aging cycles, and the health status level in an embodiment of the present invention;
FIG. 6 is a diagram of the state of health evaluation of the IGBT based on ELM in the embodiment of the invention;
FIG. 7 is a two-dimensional set of sweet spots generated by the sweet spot set rule in the embodiment of the present invention;
FIG. 8 is a Lave flight diagram of the nonlinear convergence factor in an embodiment of the present invention;
FIG. 9 is a graph comparing the BP neural network, ELM, FDA-ELM, IFDA-ELM model results in the present invention;
FIG. 10 is a graph comparing the relative error of BP neural network, ELM, FDA-ELM, IFDA-ELM models in the present invention;
FIG. 11 is a comparison graph of absolute errors of BP neural network, ELM, FDA-ELM, IFDA-ELM models in the present invention.
Detailed Description
The embodiments of the present invention will be described in detail with reference to the accompanying drawings.
The design idea of the invention is as follows: in practical engineering application, the existing RUL prediction model has the problems of high complexity, high time cost and unsatisfactory evaluation precision. The method for predicting the residual life of the IGBT is provided, and particularly the residual life of the IGBT is predicted based on an IGBT residual life prediction method of an improved runoff flow direction algorithm optimization extreme learning machine; obtaining IGBT aging test data including junction temperature data, aging cycle frequency data and saturation pressure drop data through an IGBT power cycle aging test, preprocessing the aging data of the IGBT and analyzing the correlation of the aging data and the health state of the IGBT, establishing an IGBT residual life prediction model based on an improved runoff flow direction algorithm optimization extreme learning machine, and finally bringing the aging parameters of the IGBT into verification of the effectiveness of the IGBT residual life prediction model based on the improved runoff flow direction algorithm optimization extreme learning machine, wherein the IGBT residual life prediction model has the highest accuracy. The method can be used for more accurately predicting the health state and the residual life of the IGBT, and is a new exploration of an information fusion method in the aspect of life prediction.
Based on the design concept, the invention provides a method for predicting the residual life of an IGBT, in particular to a method for predicting the residual life of an IGBT based on an improved runoff flow direction algorithm optimization extreme learning machine, and a flow chart is shown in figure 1 and comprises the following steps:
step 1, obtaining IGBT aging test data through an IGBT power cycle aging test, wherein the IGBT aging test data comprises junction temperature data, aging cycle number data and saturation voltage drop data.
The test adopts a semiconductor module power cycle aging system which comprises a test platform, a test area, a PC host, a cooling system and the like. The burn-in system contains 10 test platforms of the same specification, so the system has 10 independent burn-in test areas. The experimental steps comprise a power cycle accelerated aging test with constant shell temperature fluctuation, a single pulse measurement test and a thermal resistance measurement test. The purpose of the power cycle accelerated aging test with constant shell temperature fluctuation is to accelerate the aging process of the IGBT and reduce the time cost for acquiring the aging parameter data of the IGBT. Firstly, accelerating module aging; secondly, after the module is aged to a certain degree, the module is subjected to static discrete data measurement; finally, parameter data are obtained through a single-pulse measurement test, and steady-state thermal resistance data of the module are obtained through a thermal resistance measurement test; the experimental study subjects selected by the present invention were: model MMG75SR120 (1200V/75A) IGBT module produced by Jiangsu macro micro technology Co. The module comprises two IGBTs and two anti-parallel diodes inside. Wherein, an IGBT and an anti-parallel free-wheeling diode are in a group and are arranged on a DBC substrate. Some of the performance parameters of the module are shown in table 1 and a cross-sectional view of the module is shown in fig. 2.
TABLE 1 MMG75SR120 Module portion Performance parameter values
Table 1 Performance parameter values ofthe MMG75SR120 module
Figure BSA0000272170800000041
And 2, carrying out aging data pretreatment on the IGBT and analyzing the correlation between the IGBT and the health state of the IGBT.
The specific implementation method of the step comprises the following steps:
and 2.1, preprocessing the data of IGBT residual life prediction by using wavelet denoising.
In order to reduce the influence of noise on the saturated voltage drop data, the saturated voltage drop data of the IGBT is preprocessed by adopting a wavelet denoising method. Because the acquired saturated pressure drop data is discrete data, the saturated pressure drop is subjected to noise reduction by adopting discrete wavelet transform. Firstly, let the wavelet parent function ψ (t) ∈ F2(R) its Fourier transform is psi' (w), i.e. the wavelet series is
Figure BSA0000272170800000042
a, b ∈ R and a ≠ 0 (1)
Wherein a is a scale factor and b is a translation factor. Secondly, discretizing a and b, wherein the discretization formula of the a and the b is as follows
Figure BSA0000272170800000043
The discrete wavelet basis is obtained according to equations (1) and (2):
Figure BSA0000272170800000051
thus, discrete wavelet transform
Figure BSA0000272170800000052
Finally, the reconstruction of the signal is formulated as
Figure BSA0000272170800000053
After wavelet transform, the signal is decomposed into a low frequency signal and a high frequency signal based on the difference in signal frequency. And then, processing the threshold value of the signal to obtain an effective signal, thereby achieving the purpose of signal noise reduction.
And 2.2, aiming at the correlation problem of the aging parameters and the health state of the IGBT, the aging parameter data of the IGBT is obtained through an accelerated aging test, and the mapping relation between the aging parameter set of the IGBT and the health state grade of the IGBT is established based on the aging parameter data analysis.
According to experimental research, the IGBT crust thermal resistance value increases along with the increase of the aging cycle number, and when the aging cycle number is 7000, the increment percentage is 18.13%. Carrying out curve fitting on the aging cycle times and the incrustation thermal resistance value through a cftool box of MATLAB to obtain a fitting formula between incrustation thermal resistance R and IGBT aging cycle times cycle:
R＝0.1664×exp(2.362×10-5cycle) (6)
wherein, cycle is the aging cycle number, and R is the junction thermal resistance of the IGBT. As can be seen from the equation (6), the aging cycle number and the junction-to-crust thermal resistance are in an exponential relationship and a nonlinear relationship. The incrustation thermal resistance value and the aging cycle number are in a one-to-one mapping relation, so the aging cycle number can represent incrustation thermal resistance. Then, the functional relationship between the number of aging cycles and the incrustation resistance is denoted as R ═ h (cycle).
According to the method, the health state of the IGBT is represented by the degradation degree of the junction-crust thermal resistance, and the health state grade of the IGBT is divided through the junction-crust thermal resistance. Degradation of the junction thermal resistance is manifested as an increase in the junction thermal resistance value. Analysis shows that the IGBT junction-crust thermal resistance and the aging cycle number are in an exponential relationship, so that the relationship between the junction-crust thermal resistance and the aging cycle number needs to be linearized, and the health state grade of the IGBT is further divided. Therefore, equation (6) is linearized with a logarithmic function and equation (7) is obtained:
ln(R)＝ln0.1664+2.362×10-5×cycle (7)
wherein, cycle is aging cycle number, and R is crusting thermal resistance. When the incrustation thermal resistance value is 1.2 times the initial value, R is 0.2005. The number of aging cycles for failure of the IGBT module can be obtained by equation (7), and its value is about 7892. FIG. 3 is a diagram of the ln (R) function.
As can be seen from FIG. 3, ln (R) is a linear function, ln (R) increases linearly with the number of aging cycles, and ln (R) has function values of-1.7934, -1.6069 in the interval [0, 7892 ]. At this time, the state of health of the IGBT may be classified according to a linear standard. The invention takes 2/3 of [ -1.7934, -1.6069] as the boundary point of the IGBT health state and the sub-health state; 1.2 times of the original value of the incrustation heat resistance is taken as a boundary point of a sub-health state and a failure state. Through calculation, the aging cycle times are 5263 and 7892 respectively, the incrustation thermal resistance values are 0.1884 and 0.2005 respectively, and the incrustation thermal resistance increment percentages are 12.75% and 20% respectively, so that the health state of the IGBT is divided into three state grades of 'healthy', 'sub-healthy' and 'invalid'. The correspondence between the crusting thermal resistance interval and the health status level is shown in table 2.
TABLE 2 relationship between crusting thermal resistance interval and health status grade
Table 2 The relationship between thermal resistance interval and health status of crusts
Figure BSA0000272170800000061
In table 2, I-health state indicates that the degradation degree of the junction thermal resistance is small in this state, the IGBT has high safety and low failure probability, but transient factors such as overvoltage or overcurrent still exist, and therefore the IGBT still fails. II-sub-health state means that in this state, the IGBT junction resistance gradually degrades, and its health state is represented by a gradual increase in IGBT failure rate, so that special attention should be paid to the IGBT that has entered the sub-health state stage and replaced before it fails. The failure state indicates that the IGBT fails at the stage, and in a national economic power system, the failure of the IGBT is avoided as much as possible. In addition, IGBT health status grading facilitates the formulation and execution of IGBT maintenance decisions.
The linearized incrustation heat resistance divides the health state grade of the IGBT. Based on the relationship between the junction thermal resistance and the aging cycle number, the relationship between the aging cycle number of the IGBT and the health state grade of the IGBT is established, and the relationship chart is shown in FIG. 4.
Based on the relationship among the aging cycle number, the incrustation thermal resistance and the health state grade established in fig. 4, the corresponding relationship among the three is brought into a diagram of the change of the incrustation thermal resistance along with the aging cycle number, and a corresponding relationship diagram among the incrustation thermal resistance, the aging cycle number and the IGBT health state grade in fig. 4 is obtained.
As shown in fig. 5, the aging cycle number and the incrustation thermal resistance value of the IGBT exhibit a one-to-one correspondence relationship; the junction thermal resistance interval corresponds to the IGBT health state grade. Therefore, the IGBT crusting thermal resistance interval is converted into a corresponding aging cycle number interval, and a numerical relation between the crusting thermal resistance interval and the aging cycle number interval is established. The IGBT health can then be evaluated by predicting the number of IGBT aging cycles. The corresponding relationship among the interval of the crusting thermal resistance, the grade of the health state and the aging cycle number is shown in Table 3.
TABLE 3 correspondence table of crusting thermal resistance interval, health state grade and aging cycle number
Table 3 The table of relationship of thermal resistance increment interval of crust，health status level and the aging cycles number
Figure BSA0000272170800000062
The state of health in table 3 means that the safety of the IGBT is high at this stage. The sub-health state means that at this stage, the junction thermal resistance gradually degrades, the safety of the IGBT gradually decreases with the passage of service time, and the failure rate gradually increases, and at this time, special attention should be paid to the IGBT which has entered the sub-health state and replaced in time before it fails. The failure state indicates that the IGBT has failed at this stage. An evaluation standard of the health state of the IGBT is established based on the table 3, and the health state is healthy when the aging cycle number is between 0 and 5263; the aging cycle number is 5263 to 7892 times that the health state is sub-healthy; the number of aging cycles is greater than 7892 health states as failed.
And 3, establishing an IGBT residual life prediction model based on an improved runoff flow direction algorithm optimization extreme learning machine (IFDA-ELM).
In this step, the Extreme Learning Machine (ELM) can determine the mapping relationship between the input amount and the output amount thereof by learning the IGBT aging test data. Then, the aging parameter set of the IGBT is used as the input of the ELM, and the number of IGBT aging cycles is used as the output of the ELM. Obtaining the relation between the aging parameter set and the aging cycle times through model training; and obtaining a predicted value of the aging cycle times through testing, and accurately judging the health state grade of the IGBT according to the IGBT health state evaluation standard. The diagram of the state of health evaluation of the IGBT based on the ELM is shown in fig. 6.
The Extreme Learning Machine (ELM) based IGBT health status assessment procedure is as follows.
(1) In data driving, the IGBT aging data obtained by the power cycle accelerated aging test is selected as the driving data of the ELM. Of these, 360 (T) sets were randomly selected from 924 sets of test dataj，Vce(sat)，IcCycle) data and separating the data into training data and test dataThe volume ratio is 9: 1. The input of the model is the collector current IcJunction temperature TjAnd saturation pressure drop Vce(sat)The output of the model is the aging cycle number cycle.
(2) The test data were normalized by the normalization formula shown in equation (8). The data normalization is to perform non-dimensionalization processing on the data, so that the influence of units on different types of data is eliminated, and the data normalization processing can accelerate the learning speed of the model.
Figure BSA0000272170800000071
Where xi is the ith sample value of the original data set x, xmaxIs the maximum value, x, in the original data setminIs the minimum in the original data set, g (x)i) Is xiNormalized values.
In order to improve the searching capability and convergence performance of the runoff Flow Direction Algorithm (FDA), the method is improved from three stages of early stage, middle stage and later stage of FDA algorithm optimization. In the early stage: and (5) initializing a population. In the early stage of optimization, population individuals are uniformly distributed in a feasible region as much as possible, and then a global optimal solution is searched; and (3) a middle-stage search stage: the improved algorithm is not easy to trap in the local optimal solution for a long time, and can jump out the local optimal solution in time, thereby carrying out all-around search. And in the later stage, the step length of algorithm search is reduced, and the convergence of the algorithm is accelerated. The invention mainly improves the FDA algorithm from the following four aspects and obtains an improved runoff flow direction algorithm IFDA.
(1) The best set rule initializes the population.
When the intelligent optimization algorithm is used for searching for the optimal parameters in the solution domain, the initialized population individuals are required to be distributed in the solution domain as uniformly as possible because the information in the solution domain is rarely known. The population generated by the well-set rule can be distributed uniformly in feasible domains. Therefore, the runoff group is generated by adopting the optimal point set rule, and the specific method is as follows:
the optimal point set rule adopts a method of a circle division domain in dimGenerating a set of best points Y containing m points in the space of the dimensionm(i) The structural formula is shown in formula (9):
Ym(i)＝{[(r1×i)，(r2×i)，...，(rq×i)，...，(rdim×i)]，1≤i≤m，1≤q≤dim}(9)
in the formula, rq＝{cos(2π×q÷p)}；(rqX i) is rqFractional part of x i; p is the smallest prime number satisfying 2 xdim +3 ≦ p. The optimal point set matrix Y is obtained by equation (10):
Figure BSA0000272170800000072
wherein, a prime point represents a runoff, that is, the prime point set rule generates a runoff group comprising m runoff flows and each runoff comprises dim dimensional data. Y is1，Y2，Y3，...，YmRepresents the 1 st, 2 nd, 3 rd, m runoff, y in the runoff groupiqQ data and y on runoff individual iiq∈[LBq，UBq](1≤i≤m，1≤q≤dim)。
Fig. 7 is a distribution diagram of a two-dimensional point set generated by using the principle of the optimal point set, and it can be known from fig. 7 that the two-dimensional point set generated by using the principle of the optimal point set is uniformly distributed in the interval [ -1, 1 ]. Therefore, the optimal point set rule can be applied to population initialization of the FDA algorithm, and the optimal point set rule can uniformly distribute population individuals of the FDA algorithm in a feasible region, so that the runoff group is favorably searched for a global optimal solution.
(2) Adaptive levitational flight strategy
The adaptive Levis flight strategy is used for the runoff individual to jump out of a local optimal solution. And if the fitness value of the current runoff is equal to the fitness value of the historical optimal runoff, updating the flow direction of the runoff by adopting a Levy flight strategy, and judging according to the fitness value of the new runoff. Setting: after 10 cycles, if the optimal fitness value of the runoff does not change any more, the position of the runoff is updated by adopting an equation (11). The levy flight equation is shown in equation (12), and the levy flight diagram of the nonlinear convergence factor is shown in fig. 8.
Figure BSA0000272170800000081
Figure BSA0000272170800000082
In the formula, u and v both follow a normal distribution, β 1 is 1.5, and the calculation formula of Φ is shown in formula (13).
Figure BSA0000272170800000083
(3) Nonlinear convergence factor and gaussian variation
The expression (14) is an expression of the nonlinear convergence factor w, and the improved expression is shown in expression (15). The nonlinear convergence factor w has high change speed in the early stage and the middle stage, and the change speed state in the later stage tends to be stable, so that the algorithm tends to converge in the later stage. In addition, the runoff undergoes gaussian variation after the position is updated. The change information of the runoff flow direction is diversified through Gaussian variation, and the global optimal solution can be found conveniently.
Figure BSA0000272170800000084
Figure BSA0000272170800000085
As can be seen from fig. 8, the motion trajectory of the levey flight is irregular, and the motion range is large, which is helpful for the FDA algorithm to jump out of the local optimal solution and find the global optimal solution. The numerical value of the nonlinear convergence factor w in the early stage and the middle stage is large, namely the step length of runoff movement is large, and the algorithm is favorable for searching a global optimal solution; at the later stage, the value of w becomes smaller, that is, the step size of runoff movement becomes smaller, and the convergence of the FDA algorithm is accelerated.
And 4, substituting the aging test data of the IGBT into verification of the effectiveness of the IGBT residual life prediction model based on the improved runoff flow direction algorithm optimization extreme learning machine.
The specific implementation method of the step comprises the following steps:
and 4.1, importing the test data into an IFDA-ELM model, and testing the effectiveness of the IFDA-ELM model on the residual life prediction result of the IGBT by comparing with other models.
In order to verify the effectiveness of the IFDA-ELM model established by the invention in predicting the residual life of the IGBT, the method selects a BP neural network, an Extreme Learning Machine (ELM) and a runoff flow direction algorithm optimization extreme learning machine (FDA-ELM) model to compare with the IFDA-ELM model in predicting performance. In the aspect of data driving, 700 groups of saturation voltage drop data subjected to noise reduction processing by the IGBT module are selected as driving data of the model, the front 400 groups of data are used as training data, and the rear 300 groups of data are used as test data. And respectively carrying out 30 times of IGBT residual life prediction on the four models by using the group of data, and verifying the effectiveness and stability of the IFDA-ELM model in the aspect of IGBT residual life prediction. When a new data space is constructed, the number of sliding windows is set to 4. The parameter settings for the four predictive models are shown in table 4.
Table 4 model parameter settings
Table 4 The parameters setting of the models
Figure BSA0000272170800000091
And 4.2, analyzing the prediction result of the residual life of the IGBT based on the IFDA-ELM model.
Fig. 9 shows the results of predicting the remaining life of the IGBT according to the four models, where a relative error comparison graph of the models is shown in fig. 10, and an absolute error comparison graph of the models is shown in fig. 11.
As can be seen from fig. 9, the prediction effect of the remaining life of the IGBT of the ELM model is better than that of the BP neural network. However, at about 3080 th aging cycle, the predicted value of the ELM model changes abruptly, and the predicted value deviates far from the actual value, so the prediction error of the model at the point is large. In the prediction, the IFDA-ELM model has a better prediction effect, and the interval [3160, 3260] shows that the IFDA-ELM model has a better prediction effect than the FDA-ELM model. FIG. 10 is a graph showing the comparison of the relative error between the predicted results of the four models, the solid line and the dotted line respectively represent the IFDA-ELM and FDA-ELM models, and both curves are close to the horizontal axis with the relative error value of 0, so that the relative error values of the two models are small. The dashed line representing the ELM is largely located in stages below the dashed line representing the BP neural network, but there are still several points where the relative error values are relatively large. In general, the relative error value of the BP neural network is larger than that of the other three models. When the aging cycle number is near 3100, the relative error value of the ELM and BP neural networks exceeds 10%. Comparing fig. 10, it was found that the original saturated pressure drop curve abruptly changes around this point, and both models do not track the change of the original curve well. In addition, the relative error values of the IFDA-ELM and FDA-ELM models are also larger at this point, but are less than 10%. In fig. 11, curves representing four models, the traveling tendency of which is approximately the same as that of the curves in fig. 10, are represented. Fig. 10 and 11 show the first prediction results of model predictions, and in order to avoid the particularity of the first prediction results, the 30 prediction results of the four models are calculated by an error index formula, and the obtained error index values are recorded in table 5.
Table 5 shows the index values of the four model prediction results. Firstly, the optimal value of the model prediction result is analyzed, R2Representing the degree of fitting between the predicted curve and the actual curve, the IFDA-ELM model has the highest degree of fitting, and R thereof2The optimal value is 0.9934, which is better than the FDA-ELM model R2Optimum value of 0.9806, ELM model R2Optimum value of 0.9429 and BP neural network R2The optimum value of 0.9189. Secondly, according to analysis from the average value of the model prediction results, the average value of the IFDA-ELM model MAPE index is 1.2310, which is superior to the average value 1.9396 of the FDA-ELM model MAPE index, the average value 4.8171 of the ELM model MAPE index and the average value 9.6130 of the BP neural network MAPE index, and compared with the FDA-ELM model, the IFDA-ELM model MAPE index is reduced by 36.53%, and compared with the ELM model, the IFDA-ELM model MAPE index is reduced by 74.45%. Finally, analyzing from the standard deviation angle of the model prediction result, the standard deviation of the RMSE value of the IFDA-ELM model is 0.634, which is smaller than that of the RMSE value of the FDA-ELM model2.51, standard deviation of ELM model RMSE values 14.2, and standard deviation of BP neural network RMSE values 20.6. From the analysis of effectiveness and stability of the model for predicting the residual life of the IGBT, the effectiveness and stability of the IFDA-ELM model for predicting the residual life of the IGBT are superior to those of the comparison model, and the effectiveness and stability of the model in predicting the residual life of the IGBT module are verified.
Table 5 Module 30 index values of prediction results
Table 5 The index values of the 30 prediction results of the module
Figure BSA0000272170800000101
It should be emphasized that the embodiments described herein are illustrative rather than restrictive, and thus the present invention is not limited to the embodiments described in the detailed description, but also includes other embodiments that can be derived from the technical solutions of the present invention by those skilled in the art.

Claims (5)

1. A method for predicting the residual life of an IGBT (insulated gate bipolar translator) is disclosed, in particular to a method for predicting the residual life of an IGBT of an improved runoff flow direction algorithm optimization extreme learning machine, which is characterized in that: the method comprises the following steps:
step 1, obtaining IGBT aging test data through an IGBT power cycle aging test, wherein the IGBT aging test data comprises junction temperature data, aging cycle number data and saturation voltage drop data;
step 2, carrying out aging data pretreatment on the IGBT and analyzing the correlation of the IGBT with the health state of the IGBT;
step 3, establishing an IGBT residual life prediction model based on an improved runoff flow direction algorithm optimization extreme learning machine (IFDA-ELM);
and 4, substituting the aging test data of the IGBT into verification of the effectiveness of the IGBT residual life prediction model based on the improved runoff flow direction algorithm optimization extreme learning machine.
2. The method for predicting the residual life of the IGBT according to claim 1, wherein: the specific implementation method of the step 1 comprises the following steps: in the prediction of the residual life of the IGBT, under a certain aging condition, firstly, the aging of a module is accelerated; secondly, after the module is aged to a certain degree, the module is subjected to static discrete data measurement; and finally, acquiring parameter data through a single-pulse measurement test, acquiring steady-state thermal resistance data and junction temperature data of the module through a thermal resistance measurement test, and recording aging cycle number data.
3. The IGBT residual life prediction method based on the improved runoff flow direction algorithm optimization extreme learning machine as claimed in claim 1, characterized in that: the specific implementation method of the step 2 comprises the following steps:
step 2.1, preprocessing the data of IGBT residual life prediction by using wavelet denoising;
and 2.2, aiming at the problem of correlation between the aging parameters of the IGBT and the health state of the IGBT, the aging parameter data of the IGBT is obtained through an accelerated aging test, and the mapping relation between the aging parameter set of the IGBT and the health state grade of the IGBT is established based on the aging parameter data analysis.
4. The method for predicting the residual life of the IGBT according to claim 1, wherein: the specific implementation method of the step 3 comprises the following steps:
and 3.1, evaluating the health state of the IGBT based on the extreme learning machine. The ELM is an algorithm proposed on the basis of a single hidden layer feedforward neural network, and an Extreme Learning Machine (ELM) can determine the mapping relation between the input quantity and the output quantity of the IGBT by learning the IGBT aging test data. Then, the aging parameter set of the IGBT is used as the input of the ELM, and the number of IGBT aging cycles is used as the output of the ELM. Obtaining the relation between the aging parameter set and the aging cycle times through model training; and obtaining a predicted value of the aging cycle times through testing, and accurately judging the health state grade of the IGBT according to the IGBT health state evaluation standard.
And 3.2, because the evaluation effect of the Extreme Learning Machine (ELM) is influenced by the initial weight and the threshold value, but the improved runoff flow direction algorithm (IFDA) has the advantages of good optimizing effect, strong convergence performance and the like, the invention adopts the IFDA to search the optimal weight and the threshold value of the ELM and establishes an IFDA-ELM-based IGBT health state evaluation model.
5. The method for predicting the residual life of the IGBT according to claim 1, wherein: the specific implementation method of the step 4 comprises the following steps:
step 4.1, importing the test data into an IFDA-ELM model, testing the IFDA-ELM model, and predicting the residual life of the IGBT;
and 4.2, analyzing the prediction result of the residual life of the IGBT based on the IFDA-ELM model.
According to the method, IGBT aging test data including junction temperature data, aging cycle number data and saturation pressure drop data are obtained through an IGBT power cycle aging test, then the IGBT is subjected to aging data preprocessing and correlation analysis of the health state of the IGBT, then an IGBT residual life prediction model based on an improved runoff flow direction algorithm optimization extreme learning machine is established, finally, aging parameters of the IGBT are brought into verification of the effectiveness of the IGBT residual life prediction model based on the improved runoff flow direction algorithm optimization extreme learning machine, the most effective prediction model is obtained, and the method used in the text has a good reference effect on similar life prediction.
CN202210456063.2A 2022-04-28 2022-04-28 Method for predicting residual life of IGBT (insulated Gate Bipolar transistor) Pending CN114707423A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210456063.2A CN114707423A (en) 2022-04-28 2022-04-28 Method for predicting residual life of IGBT (insulated Gate Bipolar transistor)

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210456063.2A CN114707423A (en) 2022-04-28 2022-04-28 Method for predicting residual life of IGBT (insulated Gate Bipolar transistor)

Publications (1)

Publication Number Publication Date
CN114707423A true CN114707423A (en) 2022-07-05

Family

ID=82177413

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210456063.2A Pending CN114707423A (en) 2022-04-28 2022-04-28 Method for predicting residual life of IGBT (insulated Gate Bipolar transistor)

Country Status (1)

Country Link
CN (1) CN114707423A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116628633A (en) * 2023-07-26 2023-08-22 青岛中微创芯电子有限公司 IGBT real-time monitoring and service life prediction evaluation method

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116628633A (en) * 2023-07-26 2023-08-22 青岛中微创芯电子有限公司 IGBT real-time monitoring and service life prediction evaluation method

Similar Documents

Publication Publication Date Title
CN109842373B (en) Photovoltaic array fault diagnosis method and device based on space-time distribution characteristics
CN108053128B (en) Electric network transient stability rapid evaluation method based on ELM and TF
CN110362045B (en) Marine doubly-fed wind turbine generator fault discrimination method considering marine meteorological factors
Zhang et al. Reliability analysis of power systems integrated with high-penetration of power converters
CN109447441B (en) Transient stability risk assessment method considering uncertainty of new energy unit
CN110570122A (en) Offshore wind power plant reliability assessment method considering wind speed seasonal characteristics and current collection system element faults
CN112580471A (en) Non-invasive load identification method based on AdaBoost feature extraction and RNN model
CN112947672B (en) Maximum power point tracking method and device for photovoltaic cell
Zheng et al. Fault diagnosis of wind power converters based on compressed sensing theory and weight constrained AdaBoost-SVM
CN112288147B (en) Method for predicting insulation state of generator stator by BP-Adaboost strong predictor
Zhang et al. A hybrid forecasting system with complexity identification and improved optimization for short-term wind speed prediction
Shi et al. Study of wind turbine fault diagnosis and early warning based on SCADA data
CN112446518A (en) Method for predicting middle-and-long-term wind power electric quantity of EMD-XGboost based on EMD decomposition
CN114707423A (en) Method for predicting residual life of IGBT (insulated Gate Bipolar transistor)
CN115271253A (en) Water-wind power generation power prediction model construction method and device and storage medium
CN105024645A (en) Matrix evolution-based photovoltaic array fault location method
CN108694475B (en) Short-time-scale photovoltaic cell power generation capacity prediction method based on hybrid model
CN114595883A (en) Oil-immersed transformer residual life personalized dynamic prediction method based on meta-learning
CN114266396A (en) Transient stability discrimination method based on intelligent screening of power grid characteristics
CN117556369A (en) Power theft detection method and system for dynamically generated residual error graph convolution neural network
CN111244937B (en) Method for screening serious faults of transient voltage stability of power system
CN116307269B (en) Photovoltaic power generation power prediction method and device based on artificial intelligence
CN110610203A (en) Electric energy quality disturbance classification method based on DWT and extreme learning machine
CN116131313A (en) Explanatory analysis method for association relation between characteristic quantity and transient power angle stability
CN116341717A (en) Wind speed prediction method based on error compensation

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination