IEEE Power Electronics Magazine - March 2021 - 22

B. Composite Failure Precursor (CFP) Optimization
for SiC MOSFETs Based on Information Fusion

C. Early Degradation Prediction and Testing
Time Reduction of Batteries

This case study shows how to obtain a single optimal
health indicator, i.e., CFP in [10], based on multiple sources
of possibly relevant data. An optimal CFP expects to
achieve improved accuracy and reduced uncertainty in
RUL prediction based on a single source of failure precursor. Experimental analysis indicates that multiple physical
parameters are related to the degradation of SiC MOSFETs
[10]. In [10], a metaheuristic method, genetic programming
(GP), is applied to integrate multiple physical parameters
of SiC MOSFETs (i.e., the input capacitance Ciss, output
capacitance Coss, reverse transfer capacitance Crss, drainsource on-state resistance Rds(on), gate threshold voltage
VGS(th), and diode forward voltage VSD) to formulate a CFP.
As a symbolic programming method, GP can automatically
search for a nonlinear combination form of the physical
parameters that minimize a specific objective function, e.g.,
the prediction error in this case. The candidate combination form consists of explicit mathematical operations such
as add, minus, times, divide, logarithm, trigonometric functions, etc. More advanced operations can be added into the
combination forms to make the GP more flexible and powerful. As a result, a composite failure precursor is derived,
as shown in Figure 4. Compared to the widely used parameter Rds(on), the prediction accuracy by using the developed
composite failure precursor is improved by 35.3%, and the
associated uncertainty level is reduced by 16.3%, which
indicates the great potentials of information fusion. It is
worth mentioning that there is no physical implication of
the fusion function in Figure 4. This function is derived
from a data-driven perspective.

Figure 5 shows a case study presented in [11] on the early
degradation prediction and testing time reduction of batteries. AI tools are applied to determine the high-cycle-life
charging protocol that maximizes Lithium-ion battery
lifetime in the battery cycling experiment, as shown in Figure 5. Generally, the battery life is highly affected by the
charging protocols determined by three independent parameters CC1, CC2, and CC3. These three parameters formulate
a total of 224 charging protocols. It is very resource-consuming to determine the high-cycle-life protocol by extensively exploring the whole 224 candidates.
An early-prediction model enabled by a ML, elastic network, is applied to predict the lifetime with the first few
cycles of data. The elastic network establishes the mapping
relationship between the features extracted from the first
100 cycles and the final battery lifetime. Meanwhile, the
elastic network can automatically select the relevant features to improve the generalizability and interpretability of
the method. As a result, the final lifetime (average lifetime
905 cycles) of the batteries can be predicted by a model
trained using the data of the first 100 cycles. It suggests
that it is unnecessary to continue the experiment after
only 100 cycles for estimating the lifetime of a new battery,
which significantly saves the testing time. Subsequently,
a metaheuristic method, Bayesian optimization, is used
to optimize the design parameter space by balancing the
exploration of high uncertainty area and the exploitation of
the predicted high-cycle-life. With an active learning capability, Bayesian optimization can purposefully search for
the input space with a high-level uncertainty where there

Ciss

Coss

00
4,

0.8
0.4

VGs(th)

00
0
6,
00
0
8,
00
0
VSD

0
-0.2
0

1,000

2,000 3,000 4,000

0.5

1
0
-1
-2

CFP = sin (Rds(on)).cos(In(VGS(th)))

00
0
6,
00
0
8,
00
0

0
00
0
2,

00
0
6,
00
0
8,
00
0

0
00
0

-0.5
-1

5,000 6,000

Monitoring Time (Cycle)
Fusion Function:

Monitoring Time (Cycles)
FIG 4 Composite failure precursor based on information fusion integrating multiple data sources [10].

Device 5
Device 6
Device 7

Device 1
Device 2
Device 3
Device 4

0.2

00
0
6,
00
0
8,
00
0

0
00
0

-2

Composite Failure Precursor (CFP)

0.6

0
00
0

Rds(on)

0.5
0
-0.5
-1

0
00

00
4,

00
2,

Crss

Physical Signals

-2

IEEE POWER ELECTRONICS MAGAZINE

z March 2021

7,000

IEEE Power Electronics Magazine - March 2021

Table of Contents for the Digital Edition of IEEE Power Electronics Magazine - March 2021