IEEE Power Electronics Magazine - March 2021 - 23

Training
Dataset
Battery Cycling

Bayesian
Optimization

Early Outcome Predictor
Voltage
Cycling
Data
(From First
100 Cycles)

Capacity

Cycle Life
950

670

540

970

Cycle Life
Predictions

780

590

(From ML
Model)

730

980

Parameter 3
(CC3)

Parameter 2
(CC2)

Parameter 1
(CC1)

Modify Battery
Materials and
Processes

Recommended Charging Protocols (CC1, CC2, and CC3)

High Uncertainty
(Exploration)
High Cycle Life
(Exploitation)

FIG 5 Procedure of the AI-assisted experiment design for time-efficient Lithium-ion battery testing [11].

D. Online Anomaly Detection in DC/DC Converters
Based on Multiple Statistical Features
In [12], a case study on anomaly detection in dc-dc converters is presented. Seven statistical features are obtained in
Step 1 shown in Figure 2, from the output voltage of dc-dc
converters with unknown topology. They are the range,
mean, standard deviation, skewness, kurtosis, entropy, and
the centroid of the output voltage, as defined in [12]. Gaussian process regression (GPR) method is applied to learn the
normal behavior of a dc-dc converter for anomaly detection.

It can provide probabilistic outputs with inherent uncertainty quantification capability, as shown in Figure 6. First,
the previous 300 sampling data points of the output voltage
are applied for the model training, resulting in a model indicating the normal behavior of the converter. This model can
generate the normal voltage output, as the data points
between 300 and 500. The GPR also provides the 95% confidence interval of the normal voltage to indicate the normal
operational range, i.e., the uncertainty quantification capability. The extreme values of the seven statistical features
are used to determine the anomaly behavior. Due to the

2.38
2.37
Output Voltage (V)

is a high probability to obtain the optimal high-cycle life.
Meanwhile, this high-level uncertainty can be reduced with
more data from this space. With this recommendation, as
a closed-loop, the subsequent experiment design can specifically focus on this input space and collect more training
data to reduce the model uncertainty. In this case, the recommended charging protocols (i.e., the optimal combination of parameters 1, 2, 3 shown in Figure 5) are obtained
from the optimization results, which serve as a guide to
modify the experiment setting in the next experiment iteration. By avoiding uninformative testing, the experimental
parameter space can be optimized to reduce the number of
experiments. As a result, the high-cycle-life charging protocols are rapidly identified from 224 candidates in 16 days,
compared to more than 500 days required by a conventional
testing method. Even though this case study primarily
focuses on testing, the early degradation prediction method
could be extended to component degradation prediction in
the operation phase. It also demonstrates a time-efficient
testing method for collecting degradation data, which could
be used for condition monitoring.

2.36
2.35
2.34
2.33
2.32
2.31
0

50 100 150 200 250 300 350 400 450 500
The Number of Sampling Points
95% Confidence Interval
The Training Data
The Normal Output

FIG 6 Estimated output voltage of a dc-dc converter based on
the Gaussian process regression (GPR) model [12].

March 2021

z IEEE POWER ELECTRONICS MAGAZINE

IEEE Power Electronics Magazine - March 2021

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