20
th
Australian International Aerospace Congress, 27-28 November 2023, Melbourne
Domain-driven Residual Useful Life Estimation
Navid Zaman
1
, Daniel Chan
1
and Chris Stecki
1
1
PHM Technology, 120 Queens Parade, North Fitzroy, Victoria, 3068, Melbourne, Australia
Abstract
Predictive maintenance is widely agreed upon as the superior method for maintenance
scheduling over time/condition-based choices. This is where the remaining life cycles of each
part/component as well as the system in its entirety are forecast, informing the optimal period
to perform a maintenance action. In current literature, this is often done using a value or set of
values representing the degradation of the components or system; such a value is called a Health
Index (HI), built from a function of the sensors on the machine(s) to be monitored. This paper
will introduce and expand upon methods to utilise a flexible and domain-defined (HI), derived
from the use of a Digital Risk Twin (DRT). The twin is used to capture domain knowledge
directly from subject matter experts to determine the optimal HI at both the component stages
in addition to one’s representative of the system, thus defining at various levels of indenture.
The HI is coupled with state-of-the-art deep learning and machine learning techniques to
confidently forecast trends observed in a system or its individual parts to enable prognostics
with a high calibre of predictive integrity. Syndrome Diagnostics (SD), a tool to incorporate
the research and prototypical work laid out will also be presented.
Keywords: predictive maintenance, health index, residual useful life, digital risk twin, machine
learning, deep learning, syndrome diagnostics
Introduction
An important part of Predictive Maintenance (PdM) is the ability to reliably and accurately
forecast the remaining life of a component or system to build maintenance schedules that are
efficient. To this effect, the emerging technology of machine learning (ML) and deep learning
(DL) have been widely adopted. However, quick embrace of these methods involves risk in
form of spurious correlation and the hefty demand of information required from the domain at
various parts of the process, especially at the prediction result level, where ideally minimal
interpretation should be required.
Using a bearing system, this paper will demonstrate the use of timely collection of domain data
which informs the rest of the process and how this information is utilized in improving the
quality of ML/DL algorithm predictions. The approach focuses on understanding the problem
component/system and then using state-of-the-art technology to solve its remaining life
estimation. Much of the methodology in this paper follows the book [1] closely.
Body of the Paper
Engineering Context
Residual Useful Life is defined as the amount of time the component within a system can
operate for before it reaches a failure that requires attention in the form of maintenance actions.
Maintenance is often scheduled to refresh the life of the component; however, it is a form of
preventative maintenance and is sub-optimal as there may be significant life remaining for the
maintained component. Maintenance performed on the component once the failure has occurred
is often detrimental to the operations as it often costs the most.
To ensure the system is available at any given time, it is important to calculate the residual
useful life to accurately schedule an optimized predictive maintenance program.
