IEEE Power Electronics Magazine - March 2021 - 19

ondition monitoring is a proactive measure to
realize operation optimization, predictive
maintenance, and high availability of Power
Electronic Systems (PES). It is demanded by
reliability-, safety-, or availability-critical applications. The core of condition monitoring is a prediction
based on historical and present information. Artificial
Intelligence (AI) could play a role in addressing optimization, regression, and classification problems in predicting the operation or health status of PES. Besides AI
algorithms, quality data collection, objective formulation, and result validation require an in-depth understanding of the PES. The nexus between PES and AI
expects to create overarching effects in the condition
monitoring area. This article presents exploratory efforts
in the data-driven condition monitoring of PES in the
view of existing challenges and emerging opportunities.

Introduction

conditions, degradation level, remaining useful lifetime
(RUL), normal or abnormal status, etc.
The condition monitoring in PES has been drastically
evolved in the last decade with recent advancements in AI
[2], [3], industrial Internet-of-Things (IIoT), and data analytics. Data-driven condition monitoring solutions have been
emerging as a promising direction. They show significant
potentials and merits, e.g., improved accuracy, flexibility,
and lower requirements on hardware circuitry and deterministic models.
Cost, implementation complexity, accuracy, and impact
on existing system operation are primary factors to benchmark condition monitoring methods. Meanwhile, it should
be noted that any added hardware circuitry or software
algorithm for condition monitoring implementation brings
new risk to the system of concern, which needs to be minimized. Several challenges in data-driven condition monitoring for PES are discussed as below:
■■Limited size of data set: PES are not a data-intensive
area compared to other data-driven applications like
image recognition or natural language processing. For
example, the time-to-failure data is scarce since the
device degradation is a very long process of more than
several years in regular usage; accelerated testing experiment is resource-consuming (e.g., time, facilities, manpower) as well. Moreover, PES are usually in high-frequency operation; the required resolution and sampling
frequency of electro-thermal stresses may not be fulfilled by the existing data set for condition monitoring purposes.
■■Heterogeneous operational environment: PES are usually operating in a wide range of climatic conditions and
dynamic loading conditions. These external factors affect
the patterns of the collected data. It results in the data
with high-level variability, including system-to-system
variability, temporal variability, measurement errors, etc.
■■Limited computational platform: the computational
platform in the PES is traditionally designed for a control
purpose, which is of limited computation power to support complex data analytics. The execution time

Various measures can be applied in the design, production,
and operation phases of PES to fulfill reliability and availability requirements. Due to the limitations and uncertainties in models and data used in the design and production
phases, the risk of unexpected failure events still exists. It
calls for proactive actions that can be facilitated with condition monitoring, where the health status of PES is monitored to identify and mitigate potential risks [1]. Moreover,
the obtained in-situ operation status can also be used for
operation optimization of PES, if it is not necessarily for
preventing unexpected failure.
Condition monitoring methods can be divided into
model-based and data-driven. Model-based methods rely
mainly on the deterministic aspects of power electronic
components and systems, such as circuit operation models and degradation models. Besides these established
models, uncertainties and unexplored system principles
exist in terms of noise factors, environmental conditions,
unknown degradation mechanisms, just to name a few.
These unknown aspects make model-based methods challenging. Artificial Intelligence (AI) provides an opportunity
to overcome such a challenge. Figure 1
shows the basic concept of AI-assisted
model generation. Model-based methEstablished
ods use the inputs and established models
Physical Models
to obtain the outputs. Data-driven methUndetermined and Unexplored
Inputs
ods can apply AI algorithms and a set of
System Physics and Principles
existing input data (with or without output
data) to generate the hypothesized models,
which would be otherwise undetermined
Learning
or unexplored. Then these generated models are used to obtain the corresponding
AI Tools
outputs for new sets of input data. For the
condition monitoring of PES, the inputs
could be electrical, thermal, mechanical, or
Data
acoustic signals, and the outputs could be
estimated component parameters, stress FIG 1 AI-assisted model generation.

March 2021

Outputs

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