NON-PEER REVIEW
20
th
Australian International Aerospace Congress, 27-28 February 2023, Melbourne
20th Australian International Aerospace Congress
ISBN number: 978-1-925627-66-4
Successful Predictive Maintenance
Peter D M Brady, Emmanuel Blanchard
MathWorks Australia
Level 6, Tower 2, 475 Victoria Avenue, Chatswood, NSW 2067
Abstract
Predictive Maintenance is a topic that is fast-growing in popularity but is somewhat of a
buzzword with many different implementations. A common and simple starting point involves
the use of a “black box” sensor on a plant or machine part to capture readings. While this does
provide some results, it misses the opportunities that a full-scale implementation of Predictive
Maintenance delivers. Instead, the most successful Predictive Maintenance implementations
that we see are fully integrated with data recording systems and automatically handle model
lifecycle management. After all, as your plant operates, and loads change, the models will drift
to a point that they need to be retrained or even removed and replaced.
This talk showcases best practices and lessons learned from successful projects and high-level
walk through of the full workflow of a typical Predictive Maintenance journey. Specifically, we
begin with the raw data and early gains made with condition monitoring before concluding with
full model life cycle management. Case studies illustrate these steps.
Keywords: black box model, lifecycle management, model drift, predictive maintenance
Introduction
Our strongest lesson learnt from supporting multiple organisations to deploy active systems is
that PdM is a journey. It takes time to develop mature systems that manage the life cycle plant
assets and PdM models. Typically, we see the main components of the PdM journey as:
1. Initial data gathering.
2. Anomaly detection.
3. Condition based monitoring.
4. Static predictive maintenance models.
5. Total life cycle maturity.
Initial Data Gathering
The most important realisation during the initial data-gathering phase is that some of your
assumptions will prove incorrect. In our general PdM task work, we found that the leading
indicator of project success is the quantity and quality of the data representing the plant items
under consideration. Conceptually, the guidance is simple: gather more than you need. For
example, with rotational machinery, you generally need to sample much higher than the basic
Nyquist frequency to ensure capturing the high-speed system dynamics generating harmonics.
The main complaint with this approach is the sheer volume of data generated across an asset
fleet. However, this is just phase one. Established techniques can handle high-frequency data
capture at the edge in a mature system.
