A Relevance Vector Machine-Based Approach for
Remaining Useful Life Prediction of Power
MOSFETs
Yu Zheng
1,2,3
, Lifeng Wu
1,2,3
, Xiaojuan Li
1,2,3
and Cuixiang Yin
1,2,3
1
College of information Engineering, Capital Normal University
2
Beijing Engineering Research Center of High Reliable Embedded System, Capital Normal University
3
Beijing Key Laboratory of Electronic System Reliable Technology, Capital Normal University
Beijing, China
wooleef@gmail.com
Abstract—Accurate prediction of the RUL (remaining useful life)
of a degradation component is crucial to the PHM for an
electronic system. Power MOSFETs are widely used as essential
components of electronic and electrical subsystems and its
degradation has got more and more attention. This paper
introduces a prognostic method which based on relevance vector
machine and a degradation model to predict the RUL of power
MOSFET. The proposed method uses relevance vector machine
to find the relevance vectors. And then use relevance vectors to
find the representative vectors. The degradation model is
obtained by fitting the representative vectors. Then the RUL of
power MOSFETs can be estimated by extrapolating the
degradation model to a failure threshold. In the prediction
process, we will update the degradation model when the
difference of the predictive value and measured value exceeds the
predefined value. The results show that the proposed method can
provide better RUL estimation accuracy for power MOSFETs.
Keywords- PHM; degradation; prediction; RUL
I. INTRODUCTION
The RUL is defined as the time of an asset or a system
which can still be used before it fails. Power semiconductor
devices are critical components for electronic and electrical
subsystems. And it is more and more important as part of
power converter circuits in many aspects, such as
communications, aerospace, high-speed, radar systems, etc.
Therefore, it is critical to predict the RUL of a component or
system accurately. In this way, the operator can take
measurement timely to avoid accident.
Actual condition is the basis to estimate the health state of
components or a system. In one paper, the health state of
components was estimated according to the ring signal [1]. In
another, the degradation characteristic was extracted by an
online non-intrusive method based on Volterra series [2]. But
the degradation characteristic is not easy to obtain in these
papers. Many researchers try to find the precursor that can be
got easily by analyzing failure mechanism and aging
experiments [3-8]. And they found that the increase of junction
temperature leads to die-attach degradation, and On-resistance
will increase correspondingly. Otherwise, on-resistance can be
obtained by measurement in practice. Therefore, on-resistance
is used as the precursor of failure to predict the RUL.
To predict the RUL, there are many data-driven methods,
such as neural network, particle filter, Bayesian inference SVM
(support vector machine) and RVM (relevance vector machine).
Celaya J. R. et al. use Gaussian process regression [6] and
external Kalman filter [7] to predict RUL based on the
degradation data which got in the aging experiments. And
another method, RVM, is gaining attention to perform
prediction of RUL in many fields, such as pump impellers [9]
and lithium-ion batteries [10]. The reason why it can be widely
used is its good generalization performance and ability of
sparse inferred predictors.
Consequently, a RVM-based method is proposed to predict
the RUL of power MOSFETs dynamically in this paper. RVM
is used to train the available degradation data of power
MOSFETs in order to get the RVs (relevance vectors). The
obtained RVs are used to find the representative vectors. Then
the degradation model is proposed to fit the representative
vectors. After the degradation model is determined, RUL can
be estimated by extrapolating the degradation model to the pre-
definite failure threshold. During the prediction process, the
degradation model will be updated adaptively according to the
proposed strategy.
II. INTRODUCTION OF RVM
RVM is proposed by Michael E. Tipping in [13] and it is a
special sparse linear model. It has a similar form with SVM
[11,12] but its kernel function doesn’t need to satisfy Mercer’s
condition while SVM must meet this condition. Superior
generalization performance and shorter time for prediction is
the most compelling feature of RVM [13,14]. Besides, RVM
provides posterior probabilistic outputs.
National Natural Science Foundation of China (No.61070049, No.61202027)
National Key Technology R&D Project (No.2012DFA11340)
Beijing Natural Science Foundation of China (No.4122015)