IEEE Robotics & Automation Magazine - September 2023 - 15

PROGNOSTICS AND HEALTH MANAGEMENT SYSTEM
Data-driven prognostics and health management (PHM)
[16] is an em erging engineering discipline that facilitates
predictive maintenance scheduling. A
PHM system is expected to go far
beyond today's best practice of preventive
maintenance in the maritime
industry, which uses predetermined
maintenance intervals based on ex -
perience. Development of the twin
ship provides a unified solution
that saves different types of data of
critical machinery in a database.
These data can be then explored and
utilized to establish a data-driven
PHM system using data mining and
ML algorithms [17].
Figure 9 presents a schematic illus0
-100
-200
-300
1,400
1,450 1,500 1,550 1,600
Time (s)
(a)
500
1,650 1,700 1,750 1,800
1,400 1,450 1,500 1,550 1,600
Time (s)
(b)
tration to develop a data-driven PHM
system for the diesel engine on board
Gunnerus. Engine data include timeseries
data from installed sensors, event
data registered by ship operators, and
essential documentation provided by
manufacturers. These data can be used
by an autolabeling pipeline to generate
labeled data for ML model training. For
instance, the time-series sensor data can
be aligned with the event data to identify
normal operating, fault, and run-to-failure
data. The normal operating data are
used to train an anomaly detection algorithm,
which can be used to tell whether
the engine has deviated from normal
operating conditions. A fault classifier
is trained using the fault data, which
classify what kind of fault occurs on the
engine. The run-to-failure data are used
to train an ML regression model that
predicts the remaining useful life of an
engine before it fails completely. These
three different types of trained ML
models can therefore provide essential
information on the operating conditions,
fault types, and remaining useful life
of the engine through real-time sensor
data. Such information will be used to
inform the ship operators to develop an
optimal maintenance schedule for the
engine that eliminates failures, which is
expected to be more effective and less
costly than the preventive maintenance
currently used on board.
Currently, only anomaly detection
and remaining useful life prediction
models are developed for the diesel
200
-200
1,400 1,450 1,500 1,550 1,600
Time (s)
(c)
Twin
Ref
FIGURE 7. The validation of twin-ship docking in a simulator versus real ship in the field in
terms of positions. Ref: real ship experimental data.
Surge Speed
2
1.5
1
0.5
1,400 1,450 1,500 1,550 1,600
Time (s)
(a)
Sway Speed
0.4
0.2
-0.2
1,400 1,450 1,500 1,550 1,600
Time (s)
(b)
Twin
Ref
FIGURE 8. The validation of twin-ship docking in a simulated VS real ship in the field in
terms of speeds.
SEPTEMBER 2023 IEEE ROBOTICS & AUTOMATION MAGAZINE
15
1,650 1,700 1,750 1,800
1,650 1,700 1,750 1,800
1,650 1,700 1,750 1,800
1,650 1,700 1,750 1,800
engine on board Gunnerus. For anomaly detection, a long
short-term memory (LSTM)-based variational autoencoder
(VAE) is trained with normal operating data [18].
Velocity (m/s)
Velocity (m/s)
Heading (°)
East (m)
North (m)

IEEE Robotics & Automation Magazine - September 2023

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - September 2023