IEEE Power Electronics Magazine - March 2023 - 60

performance optimization, condition monitoring, and
lifetime extension, among others [7], [8], [9]. Establishing
the relationship between electric power losses and temperature
variation has been studied and led to the development
of advanced analytical or computational thermal
models [10]. However, precise knowledge of the electric
power losses for estimating changes in thermal behavior
is necessary. Identifying the losses during the system
operation while several unexpected factors influence its
behavior is a challenging and sometimes unsolved task
[11]. Therefore, static thermal models under known operating
conditions are usually employed in the end. In order
to overcome this issue, all factors that impact thermal
behavior must be used in a model that tracks real-time
temperature changes utilizing a formed relationship
between measured signals and temperatures [12].
An ML model can be used to establish this relationship
by understanding how already measured input parameters,
X, are correlated with output temperature signals, Y, for the
power converter, d, or even the motor, m (Figure 1). Adding
more layers to a neural network structure may further assist
in considering even more abstract phenomena without
explicitly defining their interactions with the temperature.
This article explores the utilization of existing measured
signals in today's electric drives to estimate the
case temperature of power electronics modules operating
under static and dynamic conditions. An accurate
loss estimation is not required because the ML model can
account for their impact on the temperature variation
directly from the measured signals. The case temperature
of a power electronics module has been used as an illustrative
example; nevertheless, similar results and conclusions
can be drawn for estimating the thermal behavior of
electric motors or any other powertrain component.
Specifications of the Machine-Learning Model
The development of any ML model requires the utilization of
a pipeline that assembles several steps and cross-validates
them together by setting different parameters (Figure 2). The
first step of such a pipeline is the data collection and preparation
for training the ML model. The main goal is to develop
a data-driven thermal model of a power electronics module,
so the drive must undergo several tests where speed and
load conditions are varied. For this purpose, seventeen profiles
of about eight hours each with both fast and slow
dynamics have been created. The specifications of the investigated
electric powertrain have been summarized in Table 1
where the considered converter topology is that of the classical
two-level dc-ac voltage source converter.
During the execution of each profile, several internal
drive parameters have been captured with a sampling frequency
of about 1 Hz, through a low-rate interface usually
available to customers. At this point, the cross-validation
of all the captured parameters dictates the ones that must
be added to the model because they influence the power
module temperature. The selected parameters (Table 2)
are directly related to the temperature changes, whereas
the ambient temperature is needed to provide the environmental
conditions.
More specifically, in electric drives, the power electronics
module is the main heat source due to the conduction
and switching losses. The drive current and output power,
which have been selected as model inputs, are clearly
associated with the conduction losses. The switching frequency,
which affects the switching losses, is not among
Table 1. Specifications of the investigated setup.
Motors
Test motor
Type
Power rating
Voltage rating
Current rating
Torque rating
Pole pair number
Nominal speed
Cooling means
Drives
Type
Topology
Power rating
Current rating
Cooling means
Induction machine
15 kW
400 V
30.6 A
97 N m
2
1478 rpm
Attached fan blades
Test drive
ACS880-01-038A
2-level, IGBT
26 kVA
38 A
Forced air
FIG 2 Process of model development.
60 IEEE POWER ELECTRONICS MAGAZINE z March 2023
IEEE Power Electronics Magazine - March 2023

Table of Contents for the Digital Edition of IEEE Power Electronics Magazine - March 2023
