CN113239654B - IGBT remaining life prediction method based on FIG and IPSO algorithm - Google Patents

IGBT remaining life prediction method based on FIG and IPSO algorithm Download PDF

Info

Publication number
CN113239654B
CN113239654B CN202110558620.7A CN202110558620A CN113239654B CN 113239654 B CN113239654 B CN 113239654B CN 202110558620 A CN202110558620 A CN 202110558620A CN 113239654 B CN113239654 B CN 113239654B
Authority
CN
China
Prior art keywords
particle
fuzzy
samples
data
igbt
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.)
Active
Application number
CN202110558620.7A
Other languages
Chinese (zh)
Other versions
CN113239654A (en
Inventor
王尚亭
刘震
程玉华
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.)
University of Electronic Science and Technology of China
Original Assignee
University of Electronic Science and Technology of China
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 University of Electronic Science and Technology of China filed Critical University of Electronic Science and Technology of China
Priority to CN202110558620.7A priority Critical patent/CN113239654B/en
Publication of CN113239654A publication Critical patent/CN113239654A/en
Application granted granted Critical
Publication of CN113239654B publication Critical patent/CN113239654B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/30—Circuit design
    • G06F30/36—Circuit design at the analogue level
    • G06F30/367—Design verification, e.g. using simulation, simulation program with integrated circuit emphasis [SPICE], direct methods or relaxation methods
    • 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
    • G06F30/25—Design optimisation, verification or simulation using particle-based methods
    • 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
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/004—Artificial life, i.e. computing arrangements simulating life
    • G06N3/006—Artificial life, i.e. computing arrangements simulating life based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO]
    • 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

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Computer Hardware Design (AREA)
  • Evolutionary Computation (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Geometry (AREA)
  • Artificial Intelligence (AREA)
  • Software Systems (AREA)
  • Microelectronics & Electronic Packaging (AREA)
  • Computational Linguistics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Medical Informatics (AREA)
  • Data Mining & Analysis (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses an IGBT (insulated gate bipolar transistor) residual life prediction method based on FIG (Fig) and IPSO (Internet protocol Security) algorithms, which comprises the steps of firstly obtaining collector-emitter saturation voltage drop time sequence samples of a plurality of groups of IGBTs through an accelerated IGBT life experiment and carrying out normalization processing; processing the normalized sample by using a fuzzy information granulation processing algorithm (FIG), so that the saturated voltage drop of the collector and emitter of the IGBT is converted from point prediction to interval prediction; then, a weight combination prediction model is constructed through high-order polynomial fitting, and the weight combination prediction model is optimized through an IPSO algorithm; and finally, predicting the residual service life of the IGBT in real time.

Description

IGBT remaining life prediction method based on FIG and IPSO algorithm
Technical Field
The invention belongs to the technical field of reliability analysis of power semiconductor devices, and particularly relates to a method for predicting the residual life of an IGBT (insulated gate bipolar translator) based on an FIG (field-induced breakdown) algorithm and an IPSO (Internet protocol security) algorithm.
Background
An Insulated Gate Bipolar Transistor (IGBT) is a power-level semiconductor element, can realize high-voltage and low-voltage conversion, has the characteristics of high efficiency and high switching speed, and is widely applied to key fields of national defense and military industry, new energy power generation and the like. The IGBT module plays a key role as a core part of the system, directly affecting the performance and reliability of the whole system, and the module is susceptible to the working conditions, so accurately predicting the Remaining service Life (RUL) of the IGBT is the key to optimize the service Life. Prediction techniques for the remaining service life of the IGBT module help engineers schedule maintenance times, optimize operating efficiency and avoid planning out to present a module failure.
In view of the above problems, there are currently two main solutions: a prediction technique based on degradation mechanism analysis and a prediction technique based on data driving. The prediction technology based on the degradation mechanism analysis can obtain the RUL information of the IGBT device from the material perspective, but needs to have quite deep understanding and analysis on the essential characteristics of the device manufacturing material and the process manufacturing process. Due to constraints such as complexity of an electronic system and uncertainty of an external environment of the electronic system, a physical model obtained by analyzing a degradation principle of a device is prone to lose a nonlinear relation among object parameters, and insufficient accuracy is achieved. On the other hand, the constructed model is often closely related to the specific model of the device, which is in conflict with the rapidly increasing product types of the IGBTs in the market, so that the method necessarily brings hysteresis of the prediction result. Therefore, the establishment of an accurate degradation mechanism model is complex and is rarely applied to engineering. The prediction technology based on data driving learns the mapping relation between input and output from historical aging data of the IGBT device, and establishes a nonlinear, non-transparent and non-specific object-oriented model inside the IGBT device to calculate the RUL value of the related device. The method based on data driving has strong adaptability and high real-time performance. If an accurate residual life prediction model of the IGBT can be established, the prediction precision is greatly improved.
In the big data information age, processing, analysis, and prediction of data can help us solve many problems. Machine Learning (ML) is an indispensable method for large data prediction. The Swarm Intelligence (SI) is mainly to simulate the swarm behavior of insects, animals, birds and fish. And adopting an SI algorithm to take parameters needing to be optimized as an optimization target, and taking the ML result as an optimized fitness value. Improved Particle Swarm Optimization (IPSO) is Improved on the basis of one of the common SI algorithms.
Disclosure of Invention
The invention aims to overcome the defects of the prior art and provides an IGBT (insulated gate bipolar transistor) residual life prediction method based on a FIG (Fuzzy Information Granulation) algorithm and an IPSO (Internet protocol security) algorithm.
In order to achieve the above object, the present invention provides a method for predicting remaining life of an IGBT based on FIG and IPSO algorithms, which is characterized by comprising the following steps:
(1) collecting collector-emitter saturation voltage drop time sequence samples of N groups of IGBTs;
(1.1) acquiring collector-emitter saturation voltage drop time sequence samples of N groups of IGBTs by accelerating IGBT life experiment
Figure BDA0003078078390000021
N is 1,2, …, N, and is used as a characteristic parameter of the IGBT failure process, wherein,
Figure BDA0003078078390000022
representing the M-th collector-emitter saturation voltage drop data in the n-th group of samples, wherein M represents the maximum number of collected data in the n-th group of samples;
(1.2) voltage drop time series for collector-emitter saturation
Figure BDA0003078078390000023
Carrying out normalization processing;
each will be
Figure BDA0003078078390000024
According to a single mapping function:
Figure BDA0003078078390000025
performing a normalization, wherein
Figure BDA0003078078390000026
Time sequence of voltage drop for collector-emitter saturation
Figure BDA0003078078390000027
Each of which
Figure BDA0003078078390000028
After the above normalization, each component is normalized to [0,1 ]]Number of (2) to obtain a data set Xn＝(xn1,xn2,…,xnm,…,xnM)；
(2) For data set XnCarrying out fuzzy information granulation processing (FIG);
(2.1) on dataset Xn＝(xn1,xn2,…,xnm,…,xnM) Windowing: in data set XnIn the method, every J elements are divided into a window, F is the number of the windows, F represents the total number of the windows, and F is 1,2, …, F,
Figure BDA00030780783900000211
(2.2) arranging J elements in the f-th window in an ascending order, and representing the f-th window after the ordering as
Figure BDA0003078078390000029
Representing the jth element in the f window in the nth set of samples;
(2.3) with XnfConstructing a fuzzy particle g for a domain of discourse;
Figure BDA00030780783900000210
g is a fuzzy concept and comprises three types of Low, R and Up, wherein Low represents the minimum value of the change of the fuzzy particle G, R represents the average value of the change of the fuzzy particle G, and Up represents the maximum value of the change of the fuzzy particle G;
(2.4) taking a triangular fuzzy particle form as a membership function, and depicting the fuzzy particles g through the membership function;
Figure BDA0003078078390000031
wherein a is the support lower boundary of the triangular fuzzy particle, b is the support upper boundary of the triangular fuzzy particle, and d is the nuclear parameter of the triangular fuzzy particle;
(2.5) granulating the engraved fuzzy particles g by adopting a Witold Pelrycz fuzzy granulation algorithm, and obtaining three parameter values of a, b and d in the granulation process;
Figure BDA0003078078390000032
Figure BDA0003078078390000033
Figure BDA0003078078390000034
(2.6) repeating steps (2.2) - (2.5) for each window XnfCarrying out fuzzy information granulation treatment to obtain the fuzzy information
Figure BDA0003078078390000035
And
Figure BDA0003078078390000036
(3) performing high-order polynomial fitting on the data of the three levels of Low, R and Up under the f-th window in the nth group of samples;
(3.1) setting the coefficient matrix of the fitting function under the f window in the n group of samples as
Figure BDA0003078078390000037
Represents the alpha coefficient under the f window in the n group of samples; setting the parameter matrix of the fitting function under the f window in the n group of samples as
Figure BDA0003078078390000041
Represents the second in the nth group of samplesUnder f windows the first
Figure BDA0003078078390000042
The number of the parameters is one,
Figure BDA0003078078390000043
representing the degree of the fitting polynomial, superscript T representing the transposition, G representing three fuzzy sets, including
Figure BDA0003078078390000044
Figure BDA0003078078390000045
Calculating the fitting value under the f window in the n group of samples
Figure BDA0003078078390000046
Figure BDA0003078078390000047
(3.2) adopting a least square method to carry out matrix on the coefficient under the f window in the n group of samples
Figure BDA0003078078390000048
Solving is carried out to minimize R value in the following function, and corresponding coefficient matrix is recorded
Figure BDA0003078078390000049
Figure BDA00030780783900000410
Wherein the content of the first and second substances,
Figure BDA00030780783900000411
representing the actual value of the fuzzy information granulation processing under the f window in the n group of samples;
(3.3) sampling N groups according to the method described in steps (2) to (3.2)The data matrix (X) obtained by performing high-order polynomial fitting on the data of three levels of Low, R and Up is obtainedG)*；
Figure BDA00030780783900000412
(4) Constructing a weight combination prediction model XG；
XG＝WG(XG)*+EG
Wherein, WGIs XGIn the form of a row weight matrix of
Figure BDA00030780783900000413
Line weight
Figure BDA00030780783900000414
Satisfies the following conditions:
Figure BDA00030780783900000415
and is
Figure BDA00030780783900000416
EGIs a matrix of errors between the actual values and the fitted values, in the form of
Figure BDA00030780783900000417
(5) Adopting the improved adaptive particle swarm algorithm to carry out row weight matrix WGOptimizing;
(5.1) establishing an initial population: corresponding each particle of the initial population to a weight matrix
Figure BDA00030780783900000418
Each of the elements of (a); setting the size of the population as sizepop; establishing inertia weight B and learning factor c of initial population1And c2And the maximum number of iterations is kmaxInitializing the current iteration number k to be 1;
(5.2) residual sum of squares RSS in the least squares methodIndex establishment fitness function zetaRSS；
Figure BDA0003078078390000051
(5.3) solving for min (ζ) by least squaresRSS)：
Figure BDA0003078078390000052
(5.4) updating the speed and the position of the nth particle after the kth iteration;
Figure BDA0003078078390000053
Figure BDA0003078078390000054
wherein the content of the first and second substances,
Figure BDA0003078078390000055
respectively representing the velocity and position of the nth particle after the kth iteration,
Figure BDA0003078078390000056
respectively representing the updated velocity and position of the nth particle,
Figure BDA0003078078390000057
representing the individual extreme position and the global extreme position of the nth particle at the kth generation; r is1,r2Is a random number between 0 and 1;
similarly, the speed and the position of all the particles after the k iteration are updated according to the formula;
(5.5) calculating the fitness function value after updating all the particles
Figure BDA0003078078390000058
Each will beThe updated fitness value of the particle is compared with the fitness value corresponding to the individual extreme value, if the updated fitness value of a certain particle is smaller than the fitness value corresponding to the individual extreme value, the particle is used for replacing the individual extreme value, and the corresponding optimal position is recorded; otherwise, keeping the individual extreme value unchanged; finally, selecting the individual extreme value with the minimum fitness value and the corresponding optimal position from all the individual extreme values as a global extreme value;
(5.6) checking whether the current iteration number k reaches the maximum iteration number kmaxIf yes, stopping iteration, and outputting the optimal position corresponding to the global extreme value
Figure BDA0003078078390000059
And storing the corresponding prediction model; otherwise, adding 1 to the current iteration number k, and then returning to the step (5.4)
(6) According to the optimal position
Figure BDA00030780783900000510
The corresponding prediction model carries out advanced prediction on the data group to be predicted;
(6.1) Performance failure threshold x for a given IGBTthreshold；
(6.2) collecting the collector-emitter saturation voltage drop time sequence of the IGBT to be predicted at the first t moments
Figure BDA00030780783900000511
Wherein the content of the first and second substances,
Figure BDA00030780783900000512
represents the collected p-th collector-emitter saturation voltage drop data and satisfies the following conditions:
Figure BDA0003078078390000061
p represents the number of data collected at the first t moments;
(6.3) mixing
Figure BDA0003078078390000062
Fitting the obtained data matrix according to the method described in steps (1.2) - (3.3)
Figure BDA0003078078390000063
(6.4) matching the optimal weight matrix
Figure BDA0003078078390000064
Input prediction model
Figure BDA0003078078390000065
To obtain
Figure BDA0003078078390000066
(6.5) determination
Figure BDA0003078078390000067
Whether or not to equal or exceed a performance failure threshold xthresholdIf the current time t exceeds the threshold value, recording the current time t as the critical point of the residual service life; otherwise, returning to the step (6.2) to continue to acquire data at the t +1 th moment, and processing the collector-emitter saturation voltage drop time sequence at the previous t +1 moments according to the steps (6.3) - (6.5).
The invention aims to realize the following steps:
according to the IGBT residual life prediction method based on the FIG and the IPSO algorithm, firstly, through an accelerated IGBT life experiment, collecting and emitting electrode saturation voltage drop time sequence samples of a plurality of groups of IGBTs are obtained and normalized; processing the normalized sample by using a fuzzy information granulation processing algorithm (FIG), so that the saturated voltage drop of the collector and emitter of the IGBT is converted from point prediction to interval prediction; then, a weight combination prediction model is constructed through high-order polynomial fitting, and the weight combination prediction model is optimized through an IPSO algorithm; and finally, predicting the residual service life of the IGBT in real time.
Meanwhile, the IGBT remaining life prediction method based on the FIG and the IPSO algorithm also has the following beneficial effects:
(1) the fuzzy information granulation is utilized to process the collector-emitter saturation pressure drop data, point prediction is converted into interval prediction, the influence of the characteristics of violent fluctuation and instability of the acquired time series data on the prediction is reduced, the prediction specific data is converted into the prediction data variation range, and the prediction reliability is improved.
(2) And combining the fitting results of all groups of fitting data by using different weighting strategies by using a combined weight method so as to predict the residual service life, and reducing the proportion of a sample data group with larger deviation with the test data so as to improve the prediction precision.
Drawings
FIG. 1 is a flowchart of the IGBT remaining life prediction method based on FIG and IPSO algorithm;
FIG. 2 is a schematic diagram of the IGBT's collector-emitter saturation voltage drop time series samples divided into training data and test data after fuzzy information granulation;
FIG. 3 is a three-dimensional waterfall graph of the raw data of saturation voltage drop of the 5 groups of IGBT collectors and emitters obtained from an accelerated life experiment;
FIG. 4 is a data diagram of normalized data of group 5 after fuzzy information granulation;
FIG. 5 is a result of scattering point data of IGBT collector-emitter saturation voltage drop samples to be predicted in an accelerated life experiment and a prediction curve under the IGBT remaining life prediction method based on fuzzy information granulation and IPSO algorithm provided by the invention;
fig. 6 is a three-dimensional waterfall graph comparing normalized data of all collected IGBT collector-emitter saturation voltage drops with a prediction curve, which is obtained in an accelerated life experiment.
Detailed Description
The following description of the embodiments of the present invention is provided in order to better understand the present invention for those skilled in the art with reference to the accompanying drawings. It is to be expressly noted that in the following description, a detailed description of known functions and designs will be omitted when it may obscure the subject matter of the present invention.
Examples
FIG. 1 is a flowchart of the IGBT remaining life prediction method based on FIG and IPSO algorithm;
in this embodiment, as shown in fig. 1, the method for predicting the remaining life of an IGBT based on FIG and IPSO algorithms of the present invention includes the following steps:
s1, collecting collector-emitter saturation voltage drop time sequence samples of the N-5 groups of IGBTs;
s1.1, through the accelerated IGBT lifetime experiment, as shown in fig. 1, acquiring collector-emitter saturation voltage drop time series samples of N-5 groups of IGBTs
Figure BDA0003078078390000071
n is 1,2, …,5, and is used as a characteristic parameter of the IGBT failure process, wherein,
Figure BDA0003078078390000072
representing the M-th collector-emitter saturation voltage drop data in the n-th group of samples, wherein M represents the number of collected data in the n-th group of samples; in this embodiment, taking the group 1 sample as an example, the number M of the collected data is 2232;
s1.2, collector-emitter saturation voltage drop time sequence
Figure BDA0003078078390000073
Carrying out normalization processing;
each will be
Figure BDA0003078078390000074
According to a single mapping function:
Figure BDA0003078078390000075
performing a normalization, wherein
Figure BDA0003078078390000076
Time sequence of voltage drop for collector-emitter saturation
Figure BDA0003078078390000077
Each of which
Figure BDA0003078078390000078
After the above normalization, each component is normalized to [0,1 ]]Number of (2) to obtain a data set Xn＝(xn1,xn2,…,xnm,…,xnM)；
S2, data set XnCarrying out fuzzy information granulation processing (FIG);
s2.1, data set Xn＝(xn1,xn2,…,xnm,…,xnM) Windowing: in data set XnIn the interval J, 5 elements are divided into a window, F is the number of the windows, F represents the total number of the windows, F is 1,2, …, F,
Figure BDA0003078078390000081
in this embodiment, taking the nth-1 group of samples as an example, the total number of windows is
Figure BDA0003078078390000082
S2.2, arranging 5 elements in the f-th window in an ascending order, and representing the f-th window after the f-th window is arranged in the ascending order
Figure BDA0003078078390000083
Representing the jth element in the f window in the nth set of samples;
s2.3, with XnfConstructing a fuzzy particle g for a domain of discourse;
Figure BDA0003078078390000084
g is a fuzzy concept and comprises three types of Low, R and Up, wherein Low represents the minimum value of the change of the fuzzy particle G, R represents the average value of the change of the fuzzy particle G, and Up represents the maximum value of the change of the fuzzy particle G;
s2.4, taking a triangular fuzzy particle form as a membership function, and depicting the fuzzy particles g through the membership function;
Figure BDA0003078078390000085
wherein a is the support lower boundary of the triangular fuzzy particle, b is the support upper boundary of the triangular fuzzy particle, and d is the nuclear parameter of the triangular fuzzy particle;
s2.5, granulating the engraved fuzzy particles g by adopting a Witold Pelrycz fuzzy granulation algorithm, and obtaining three parameter values of a, b and d in the granulation process;
Figure BDA0003078078390000086
Figure BDA0003078078390000087
Figure BDA0003078078390000088
s2.6, repeating the steps S2.2-S2.5, and aiming at each window XnfCarrying out fuzzy information granulation treatment to obtain the fuzzy information
Figure BDA0003078078390000091
And
Figure BDA0003078078390000092
s3, performing high-order polynomial fitting on the data of three levels of Low, R and Up in the f-th window in the n-th group of samples;
s3.1, setting a coefficient matrix of a fitting function under the f window in the n group of samples as
Figure BDA0003078078390000093
Represents the alpha coefficient under the f window in the n group of samples; setting the parameter matrix of the fitting function under the f window in the n group of samples as
Figure BDA0003078078390000094
Representing the f-th window in the n-th set of samples
Figure BDA0003078078390000095
A parameter，
Figure BDA0003078078390000096
Representing the degree of the fitting polynomial, superscript T representing the transposition, G representing three fuzzy sets, including
Figure BDA0003078078390000097
Figure BDA0003078078390000098
Calculating the fitting value under the f window in the n group of samples
Figure BDA0003078078390000099
Figure BDA00030780783900000910
S3.2, adopting a least square method to carry out coefficient matrix under the f window in the n group of samples
Figure BDA00030780783900000911
Solving is carried out to minimize R value in the following function, and corresponding coefficient matrix is recorded
Figure BDA00030780783900000912
Figure BDA00030780783900000913
Wherein the content of the first and second substances,
Figure BDA00030780783900000914
representing the actual value of the fuzzy information granulation processing under the f window in the n group of samples;
s3.3, and carrying out high-order polynomial fitting on the data of the N groups of samples at three levels of Low, R and Up according to the method of the steps S2-S3.2 to obtain a data matrix (X)G)*；
Figure BDA00030780783900000915
S4, constructing weight combination prediction model XG；
XG＝WG(XG)*+EG
Wherein, WGIs XGIn the form of a row weight matrix of
Figure BDA0003078078390000101
Line weight
Figure BDA0003078078390000102
Satisfies the following conditions:
Figure BDA0003078078390000103
EGis a matrix of errors between the actual values and the fitted values, in the form of
Figure BDA0003078078390000104
S5, adopting the improved adaptive particle swarm optimization to carry out row weight matrix WGOptimizing;
s5.1, establishing an initial population: corresponding each particle of the initial population to a weight matrix
Figure BDA0003078078390000105
Each of the elements of (a); setting the size of the population as sizepop; constructing inertia weight B and learning factor c of initial population1And c2And the maximum number of iterations is kmax100, and initializing the current iteration number k to 1;
in this embodiment, the inertia weight B keeps the inertia of the particle moving, so that it has a tendency to expand the search space. A proper B value can effectively give consideration to the search precision and the search speed, the global search and the local search, and in order to ensure the performance of the algorithm, the inertia weight B is adjusted according to the following formula:
Figure BDA0003078078390000106
wherein, Bmin＝0.8,Bmax＝1.5；
The learning factor is two parameters which represent the ability of the particle to optimally learn from the self history and optimally learn from the group history. Compared with the traditional classical learning factor value, the method can not only ensure the learning capability of the particles in the initial stage of iteration, but also strengthen the global search; the learning factor is adjusted through different evolution speeds of the particles, so that the particles adjust the learning mode by combining self conditions. Learning factor c1Representing the ability of the particle to approach the optimal potential of the individual, giving the particle a self-summarizing ability, a learning factor c2Representing the ability of a particle to approach a global optimum position, larger c early in the particle flight search1And smaller c2The particles can be better flown to the global optimal solution; in the later stages of particle flight search, the smaller c1And greater c2The particles can be made to fly in locally optimal directions, which is advantageous for accelerating the convergence of the particles. The learning factor adjustment strategy is as follows:
Figure BDA0003078078390000107
Figure BDA0003078078390000108
wherein, c1max＝2.5,c1min＝1.5,c2max＝2.5,c2min＝1.5；
S5.2, searching for the optimal function matching with the change trend of the test data by introducing an important index Residual Sum of Squares (RSS) in the least square method, and taking the index as a factor with weak evaluation adaptability. The smaller the fitness value is, the higher the optimization performance is; then, a fitness function zeta is established by utilizing the residual square sum RSS index in the least square methodRSSComprises the following steps:
Figure BDA0003078078390000111
s5.3, solving min (zeta) through a least square methodRSS)：
Figure BDA0003078078390000112
S5.4, updating the speed and the position of the nth particle after the kth iteration;
Figure BDA0003078078390000113
the updating of the particle position is related to the current position and the flight speed, when the particle is in a bad position and the flight speed is not ideal, the updated particle has poor adaptability, and the particle is guided to fly to the global optimal position based on a position updating formula; the particle position update adjustment formula is as follows:
Figure BDA0003078078390000114
wherein the content of the first and second substances,
Figure BDA0003078078390000115
respectively representing the velocity and position of the nth particle after the kth iteration,
Figure BDA0003078078390000116
respectively representing the updated velocity and position of the nth particle,
Figure BDA0003078078390000117
representing the individual extreme position and the global extreme position of the nth particle at the kth generation; r is1,r2Is a random number between 0 and 1, and the values are all 0.5;
similarly, the speed and the position of all the particles after the k iteration are updated according to the formula;
s5.5, calculating the fitness function value after updating all the particles
Figure BDA0003078078390000118
Comparing the updated fitness value of each particle with the fitness value corresponding to the individual extreme value, if the updated fitness value of a certain particle is smaller than the fitness value corresponding to the individual extreme value, replacing the individual extreme value with the particle, and recording the corresponding optimal position; otherwise, keeping the individual extreme value unchanged; finally, selecting the individual extreme value with the minimum fitness value and the corresponding optimal position from all the individual extreme values as a global extreme value;
s5.6, checking whether the current iteration number k reaches the maximum iteration number kmaxIf the position is 100, stopping iteration and outputting the optimal position corresponding to the global extreme value
Figure BDA0003078078390000121
And storing the corresponding prediction model; otherwise, adding 1 to the current iteration number k, and then returning to the step S5.4;
s6, according to the best position
Figure BDA0003078078390000122
The corresponding prediction model carries out advanced prediction on the data group to be predicted;
s6.1, setting performance failure threshold x of IGBTthreshold＝0.15；
S6.2, collecting the collector-emitter saturation voltage drop time sequence of the IGBT to be predicted at the first t moments
Figure BDA0003078078390000123
Wherein the content of the first and second substances,
Figure BDA0003078078390000124
represents the collected p-th collector-emitter saturation voltage drop data and satisfies the following conditions:
Figure BDA0003078078390000125
p represents the number of data acquired at the previous t moments, and at this time, P is 1905;
s6.3, mixing
Figure BDA0003078078390000126
Fitting the obtained data matrix according to the method described in steps S1.2-S3.3
Figure BDA0003078078390000127
S6.4, matching the optimal weight matrix
Figure BDA0003078078390000128
And (X)G)*Input prediction model
Figure BDA0003078078390000129
To obtain
Figure BDA00030780783900001210
S6.5, judging
Figure BDA00030780783900001211
Whether or not to equal or exceed a performance failure threshold xthresholdIf the current time t exceeds the threshold value, recording the current time t as the critical point of the residual service life; otherwise, returning to the step S6.2 to continuously acquire the data at the t +1 th moment, and processing the collector-emitter saturation voltage drop time sequence at the previous t +1 moments according to the steps S6.3-S6.5.
In this embodiment, to illustrate the technical effect of the present invention, a total of 6 groups of IGBT accelerated life experiments are completed for verification, wherein as shown in fig. 2, the first 5 groups of data are used for training a prediction model, and the 6 th group of data are used for real-time prediction. Fig. 3 is a three-dimensional waterfall graph of raw data of saturation voltage drops of 5 groups of IGBTs obtained by experiments. By observing the original data shown in fig. 3, the data can be found to have the characteristics of redundancy and high dimension. If the method is directly applied to life prediction, the efficiency and accuracy of prediction can be greatly reduced. Therefore, according to the actual needs and problem characteristics, in this embodiment, the fuzzy information granulation process is performed on the processed data set obtained in step S1, the fuzzy process of the present invention uses the triangular fuzzy particle form a (x, a, d, b) to determine G as the membership function, and 3 types of fuzzy granulated particles that obtain the same data sample are: low, R and Up to accurately predict the lifetime of the IGBT, and fig. 4 shows a data scattergram situation after fuzzy information granulation is performed on the 5 th group of data after normalization processing.
In order to select the optimal weight parameter in the weight combination prediction model to improve the prediction precision of the data to be predicted, an improved particle swarm algorithm is adopted to select the optimal weight parameter. In order to improve the fitting accuracy and guarantee the predicted undistorted condition, the polynomial fitting order is 9, the size of the population scale is set to be 200, and the maximum evolutionary number is 100. The design utilizes random numbers to establish an initial population.
Under the above parameter conditions, in order to better verify the prediction performance of the method proposed by the present invention, two experiments under different conditions were performed as follows: (1) testing the influence of different failure threshold APTs on the prediction accuracy of the residual life of the IGBT on the same training data point number and verification data point number, and analyzing the influence of different failure threshold APTs on the residual life of the IGBT; (2) by utilizing the running and timing functions of Matlab software, the prediction performance of the algorithm under different sizepop values of the algorithm provided by the invention is compared from two aspects of prediction precision and time consumption.
Table 1 shows the predicted results of the algorithm of the present invention in experiment (1), and Table 2 shows the predicted results in experiment (2).
TABLE 1
Failure threshold APT Actual value Prediction value Prediction error Relative error
0.15 2101 2120 19 0.904％
0.16 2108 2135 27 1.280％
0.17 2156 2155 1 0.046％
0.18 2169 2170 1 0.046％
0.19 2179 2178 1 0.046％
0.20 2194 2198 4 0.182％
In order to verify the superior performance of the algorithm of the model, the average absolute percentage Error (MSE) is calculated between the predicted data and the actual data in the test set. The MSE ranges from [0, + ∞), and an MSE of 0% indicates a perfect model, i.e., when the predicted value matches the true value completely.
Figure BDA0003078078390000131
TABLE 2
Figure BDA0003078078390000132
Figure BDA0003078078390000141
From table 1, it can be seen that when different failure thresholds APT are set, the residual life of the IGBT predicted based on fuzzy information granulation and the IPSO algorithm provided by the present invention has a relative error not exceeding 1.5%. In addition, fig. 5 (solid line part) shows a result graph of the average level prediction condition of the IGBT data of the data group to be predicted, which is predicted by the method provided by the present invention, and fig. 6 also shows a three-dimensional waterfall graph of a curve of the training data and the predicted average level R, in order to visually compare with the training data and to more visually recognize the result. By combining the table 1, the fig. 5 and the fig. 6, the method provided by the invention has the advantages that the overall trend of the IGBT experimental data can be well predicted, and the residual life can be more accurately predicted.
Besides, table 2 lists the time consumption of predicting the saturation voltage drop data of the IGBT collector and emitter in the group 6 of the accelerated life experiment by using sizepop with three population sizes, and it can be seen that when the algorithm sets different sizepop parameters, the root mean square error of the verification data is within 10%. Based on MAPE, we can judge the model is good.
In summary, the above experimental results show that the method can obtain a higher prediction progress under the condition of a small number of training samples on the basis of ensuring the operation efficiency.
Although illustrative embodiments of the present invention have been described above to facilitate the understanding of the present invention by those skilled in the art, it should be understood that the present invention is not limited to the scope of the embodiments, and various changes may be made apparent to those skilled in the art as long as they are within the spirit and scope of the present invention as defined and defined by the appended claims, and all matters of the invention which utilize the inventive concepts are protected.

Claims (3)

1. A prediction method for the residual life of an IGBT based on FIG and IPSO algorithm is characterized by comprising the following steps:
(1) collecting collector-emitter saturation voltage drop time sequence samples of N groups of IGBTs;
(1.1) acquiring collector-emitter saturation voltage drop time sequence samples of N groups of IGBTs by accelerating IGBT life experiment
Figure FDA0003479843360000011
And is used as a characteristic parameter of the IGBT failure process, wherein,
Figure FDA0003479843360000012
representing the M-th collector-emitter saturation voltage drop data in the n-th group of samples, wherein M represents the maximum number of collected data in the n-th group of samples;
(1.2) voltage drop time series for collector-emitter saturation
Figure FDA0003479843360000013
Carrying out normalization processing;
each will be
Figure FDA0003479843360000014
According to a single mapping function:
Figure FDA0003479843360000015
a normalization is performed, wherein,
Figure FDA0003479843360000016
time sequence of voltage drop for collector-emitter saturation
Figure FDA0003479843360000017
Each of which
Figure FDA0003479843360000018
After the above normalization, each component is normalized to [0 ],1]Number of (2) to obtain a data set Xn＝(xn1,xn2,…,xnm,…,xnM)；
(2) For data set XnCarrying out fuzzy information granulation processing (FIG);
(2.1) on dataset Xn＝(xn1,xn2,…,xnm,…,xnM) Windowing: in data set XnIn the method, J elements are divided into a window at intervals, F is the number of the windows, and F represents the total number of the windows
Figure FDA0003479843360000019
(2.2) arranging J elements in the f-th window in an ascending order, and representing the f-th window after the ordering as
Figure FDA00034798433600000110
Figure FDA00034798433600000111
Representing the jth element in the f window in the nth set of samples;
(2.3) with XnfConstructing a fuzzy particle g for a domain of discourse;
Figure FDA00034798433600000112
g is a fuzzy concept and comprises three types of Low, R and Up, wherein Low represents the minimum value of the change of the fuzzy particle G, R represents the average value of the change of the fuzzy particle G, and Up represents the maximum value of the change of the fuzzy particle G;
(2.4) taking a triangular fuzzy particle form as a membership function, and depicting the fuzzy particles g through the membership function;
Figure FDA0003479843360000021
wherein a is the support lower boundary of the triangular fuzzy particle, b is the support upper boundary of the triangular fuzzy particle, and d is the nuclear parameter of the triangular fuzzy particle;
(2.5) granulating the engraved fuzzy particles g by adopting a Witold Pelrycz fuzzy granulation algorithm, and obtaining three parameter values of a, b and d in the granulation process;
Figure FDA0003479843360000022
Figure FDA0003479843360000023
Figure FDA0003479843360000024
；
(2.6) repeating steps (2.2) - (2.5) for each window XnfCarrying out fuzzy information granulation treatment to obtain the fuzzy information
Figure FDA0003479843360000025
Figure FDA0003479843360000026
And
Figure FDA0003479843360000027
(3) carrying out high-order polynomial fitting on the data of three levels of Low, R and Up under each window in the nth group of samples;
(3.1) setting the coefficient matrix of the fitting function under the f window in the n group of samples as
Figure FDA0003479843360000028
Figure FDA0003479843360000029
Represents the alpha coefficient under the f window in the n group of samples; setting the parameter matrix of the fitting function under the f window in the n group of samples as
Figure FDA00034798433600000210
Figure FDA00034798433600000211
Representing the f-th window in the n-th set of samples
Figure FDA00034798433600000212
The number of the parameters is one,
Figure FDA00034798433600000213
representing the degree of the fitting polynomial, superscript T representing the transposition, G representing three fuzzy sets, including
Figure FDA00034798433600000214
Figure FDA00034798433600000215
Calculating the fitting value under the f window in the n group of samples
Figure FDA0003479843360000031
Figure FDA0003479843360000032
；
(3.2) adopting a least square method to carry out matrix on the coefficient under the f window in the n group of samples
Figure FDA0003479843360000033
Solving is carried out to minimize R value in the following function, and corresponding coefficient matrix is recorded
Figure FDA0003479843360000034
Figure FDA0003479843360000035
Wherein the content of the first and second substances,
Figure FDA0003479843360000036
representing the actual value of the fuzzy information granulation processing under the f window in the n group of samples;
(3.3) performing high-order polynomial fitting on the data of the N groups of samples at three levels of Low, R and Up according to the method of the steps (2) to (3.2) to obtain a data matrix (X)G)*；
Figure FDA0003479843360000037
；
(4) Constructing a weight combination prediction model XG；
XG＝WG(XG)*+EG
Wherein, WGIs XGIn the form of a row weight matrix of
Figure FDA0003479843360000038
Line weight
Figure FDA0003479843360000039
Satisfies the following conditions:
Figure FDA00034798433600000310
and is
Figure FDA00034798433600000311
EGIs a matrix of errors between the actual values and the fitted values, in the form of
Figure FDA00034798433600000312
(5) Adopting the improved adaptive particle swarm algorithm to carry out row weight matrix WGOptimizing;
(5.1) establishing an initial population: corresponding each particle of the initial population to a weight matrix
Figure FDA00034798433600000313
Each of the elements of (a); setting the size of the population as sizepop; constructing inertia weight B and learning factor c of initial population1And c2And the maximum number of iterations is kmaxInitializing the current iteration number k to be 1;
(5.2) establishing a fitness function zeta by utilizing residual square sum RSS indexes in the least square methodRSS；
Figure FDA0003479843360000041
；
(5.3) solving for min (ζ) by least squaresRSS)：
Figure FDA0003479843360000042
；
(5.4) updating the speed and the position of the nth particle after the kth iteration;
Figure FDA0003479843360000043
Figure FDA0003479843360000044
wherein the content of the first and second substances,
Figure FDA0003479843360000045
respectively representing the velocity and position of the nth particle after the kth iteration,
Figure FDA0003479843360000046
respectively representing the updated velocity and position of the nth particle,
Figure FDA0003479843360000047
representing the individual extreme position and the global extreme position of the nth particle at the kth generation; r is1,r2Is a random number between 0 and 1;
similarly, the speed and the position of all the particles after the k iteration are updated according to the formula;
(5.5) calculating the fitness function value after updating all the particles
Figure FDA0003479843360000048
Comparing the updated fitness value of each particle with the fitness value corresponding to the individual extreme value, if the updated fitness value of a certain particle is smaller than the fitness value corresponding to the individual extreme value, replacing the individual extreme value with the particle, and recording the corresponding optimal position; otherwise, keeping the individual extreme value unchanged; finally, selecting the individual extreme value with the minimum fitness value and the corresponding optimal position from all the individual extreme values as a global extreme value;
(5.6) checking whether the current iteration number k reaches the maximum iteration number kmaxIf yes, stopping iteration, and outputting the optimal position corresponding to the global extreme value
Figure FDA0003479843360000049
And storing the corresponding prediction model; otherwise, adding 1 to the current iteration number k, and then returning to the step (5.4);
(6) according to the optimal position
Figure FDA00034798433600000410
The corresponding prediction model carries out advanced prediction on the data group to be predicted;
(6.1) Performance failure threshold x for a given IGBTthreshold；
(6.2) collecting the collector-emitter saturation voltage drop time sequence of the IGBT to be predicted at the first t moments
Figure FDA00034798433600000411
Wherein the content of the first and second substances,
Figure FDA00034798433600000412
represents the collected p-th collector-emitter saturation voltage drop data and satisfies the following conditions:
Figure FDA00034798433600000413
p represents the number of data collected at the first t moments;
(6.3) mixing
Figure FDA00034798433600000414
Fitting the obtained data matrix according to the method described in steps (1.2) - (3.3)
Figure FDA00034798433600000415
(6.4) matching the optimal weight matrix
Figure FDA0003479843360000051
And (X)G)*Input prediction model
Figure FDA0003479843360000052
To obtain
Figure FDA0003479843360000053
(6.5) determination
Figure FDA0003479843360000054
Whether or not to equal or exceed a performance failure threshold xthresholdIf the current time t exceeds the threshold value, recording the current time t as the critical point of the residual service life; otherwise, returning to the step (6.2) to continue to acquire data at the t +1 th moment, and processing the collector-emitter saturation voltage drop time sequence at the previous t +1 moments according to the steps (6.3) - (6.5).
2. The method for predicting the remaining life of the IGBT based on the FIG and IPSO algorithm according to claim 1, wherein the inertia weight B satisfies the following condition:
Figure FDA0003479843360000055
wherein, Bmax,BminGiven the inertia weight maximum and minimum values, respectively, λ is a constant.
3. The IGBT remaining life prediction method based on FIG and IPSO algorithm of claim 1, wherein the learning factor c1、c2Satisfies the following conditions:
Figure FDA0003479843360000056
Figure FDA0003479843360000057
wherein, c1maxRepresents a learning factor c1Upper limit of adjustment of c1minThen represents the learning factor c1The lower limit of adjustment of (2); c. C2maxRepresents a learning factor c2Lower limit of adjustment of c2minThen representLearning factor c2And adjusting the lower limit.
CN202110558620.7A 2021-05-21 2021-05-21 IGBT remaining life prediction method based on FIG and IPSO algorithm Active CN113239654B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110558620.7A CN113239654B (en) 2021-05-21 2021-05-21 IGBT remaining life prediction method based on FIG and IPSO algorithm

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110558620.7A CN113239654B (en) 2021-05-21 2021-05-21 IGBT remaining life prediction method based on FIG and IPSO algorithm

Publications (2)

Publication Number Publication Date
CN113239654A CN113239654A (en) 2021-08-10
CN113239654B true CN113239654B (en) 2022-03-25

Family

ID=77138158

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110558620.7A Active CN113239654B (en) 2021-05-21 2021-05-21 IGBT remaining life prediction method based on FIG and IPSO algorithm

Country Status (1)

Country Link
CN (1) CN113239654B (en)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114859231B (en) * 2022-04-27 2023-06-09 电子科技大学 Battery remaining life prediction method based on wiener process and extreme learning machine
CN116205190B (en) * 2023-05-05 2023-08-18 湖南大学 IGBT physical model parameter acquisition method and device

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105590023A (en) * 2015-12-08 2016-05-18 三峡大学 Fuzzy granulation prediction method of performance degradation of rolling bearing on the basis of information entropy
CN109284543A (en) * 2018-09-04 2019-01-29 河北工业大学 IGBT method for predicting residual useful life based on optimal scale Gaussian process model
CN109472397A (en) * 2018-10-19 2019-03-15 东华大学 Polymerization technique parameter adjusting method based on viscosity change
CN112213660A (en) * 2020-08-25 2021-01-12 广西电网有限责任公司南宁供电局 Method for predicting residual life of power electronic device in UPS system
WO2021092845A1 (en) * 2019-11-14 2021-05-20 Elekta (Shanghai) Technology Co., Ltd. Predictive maintenance of dynamic leaf guide based on deep learning

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105590023A (en) * 2015-12-08 2016-05-18 三峡大学 Fuzzy granulation prediction method of performance degradation of rolling bearing on the basis of information entropy
CN109284543A (en) * 2018-09-04 2019-01-29 河北工业大学 IGBT method for predicting residual useful life based on optimal scale Gaussian process model
CN109472397A (en) * 2018-10-19 2019-03-15 东华大学 Polymerization technique parameter adjusting method based on viscosity change
WO2021092845A1 (en) * 2019-11-14 2021-05-20 Elekta (Shanghai) Technology Co., Ltd. Predictive maintenance of dynamic leaf guide based on deep learning
CN112213660A (en) * 2020-08-25 2021-01-12 广西电网有限责任公司南宁供电局 Method for predicting residual life of power electronic device in UPS system

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
Remaining Useful Life Estimation of Insulated Gate Biploar Transistors(IGBTs) Based on a Novel Volterra k-Nearest Neighbor OPtimally PRUNED Extreme Learning Machine(VKOPP) Model Using Degradation Data;Zhen liu等;《sensors》;20171103;第17卷(第11期);1-23 *
基于改进混沌粒子群算法的IGBT参数提取及有效性验证研究;曾春燕;《中国优秀博硕士学位论文全文数据库(硕士)信息科技辑》;20210115(第1期);I135-265 *
基于粒子群优化算法的航空发动机故障诊断与性能参数预测研究;郑波;《中国优秀博硕士学位论文全文数据库(博士)工程科技Ⅱ辑》;20180915(第9期);C031-4 *

Also Published As

Publication number Publication date
CN113239654A (en) 2021-08-10

Similar Documents

Publication Publication Date Title
CN113239654B (en) IGBT remaining life prediction method based on FIG and IPSO algorithm
CN108596212B (en) Transformer fault diagnosis method based on improved cuckoo search optimization neural network
CN109931678B (en) Air conditioner fault diagnosis method based on deep learning LSTM
CN105160444B (en) Electrical equipment failure rate determining method and system
CN107688825B (en) Improved integrated weighted extreme learning machine sewage treatment fault diagnosis method
CN110232416B (en) Equipment fault prediction method based on HSMM-SVM
CN107506865B (en) Load prediction method and system based on LSSVM optimization
CN112487702A (en) Method for predicting remaining service life of lithium ion battery
CN105426955A (en) Disturbance-based elite reverse learning particle swarm optimization implementation method
CN109543741A (en) A kind of FCM algorithm optimization method based on improvement artificial bee colony
CN110987436B (en) Bearing fault diagnosis method based on excitation mechanism
CN112686372A (en) Product performance prediction method based on depth residual GRU neural network
CN112734002B (en) Service life prediction method based on data layer and model layer joint transfer learning
CN110929451A (en) Fuel cell single voltage consistency prediction method
CN111292124A (en) Water demand prediction method based on optimized combined neural network
CN115689070B (en) Energy prediction method for optimizing BP neural network model based on monarch butterfly algorithm
CN113012766A (en) Self-adaptive soft measurement modeling method based on online selective integration
Huang et al. Research on transformer fault diagnosis method based on GWO optimized hybrid kernel extreme learning machine
Shang et al. Research on intelligent pest prediction of based on improved artificial neural network
CN116542382A (en) Sewage treatment dissolved oxygen concentration prediction method based on mixed optimization algorithm
CN115062528A (en) Prediction method for industrial process time sequence data
CN113919221A (en) Fan load prediction and analysis method and device based on BP neural network and storage medium
Bi et al. Self-adaptive Teaching-learning-based Optimizer with Improved RBF and Sparse Autoencoder for Complex Optimization Problems
CN110276478B (en) Short-term wind power prediction method based on segmented ant colony algorithm optimization SVM
CN116796406A (en) Building energy consumption optimization method based on artificial neural network and BIM

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
GR01 Patent grant
GR01 Patent grant