NON-PEER REVIEW
20
th
Australian International Aerospace Congress, 27-28 February 2023, Melbourne
Practical Predictive Maintenance Workflows
Peter Brady and Emmanuel Blanchard
MathWorks Australia
Level 6, Tower 2, 475 Victoria Avenue, Chatswood, NSW 2067
Abstract
This session will focus on a deep dive of the data cleansing, condition monitoring, and model
development phases. To illustrate, we will be using two technical data sets: (1) turbine run to
failure data and (2) wind turbine main bearing prognosis. We examine data cleansing
methodologies and the different numerical techniques to predict remaining useful life using
survival, degradation, or similarity models, depending on your system data.
While a full predictive maintenance solution to estimate remaining useful life requires data,
which takes time to gather and process, early gains can be achieved through condition
monitoring. We will spend some time dissecting this approach and how to use condition
monitoring algorithms to develop traffic light dashboards as a precursor to predictive
maintenance maturity.
Keywords: black box model, lifecycle management, model drift, predictive maintenance.
Introduction
This paper walks through a deep dive of a standard predictive maintenance workflow, which
nominally follows the following steps:
1. data import and exploration
2. feature extraction and postprocessing
3. feature importance ranking and fusion
4. model fitting and prediction
5. performance analysis.
This relates the central “dashed” region of the block diagram shown in Figure 1.
Figure 1 – Block diagram of a typical data driven predictive maintenance workflow.
It is important to note that this paper will not discuss the final, critical steps of deployment and
model drift maintenance of a predictive maintenance. Instead, it will focus on the initial
development of a model. Conceptually, drift detection is merely an extension of this training
workflow, which carries on throughout the plant’s useful life, somewhat like a continuous
improvement quality assurance process.
The data set that this deep dive is based on was collected from a 2 MW wind turbine high-speed
shaft driven by a 20-tooth pinion gear [1]. An interesting facet of this data set is that it shows a
very good use case for high-fidelity analysis on the turbine rather than streaming and central
