20
th
Australian International Aerospace Congress, 27 February - 1 March 2023, Melbourne
20th Australian International Aerospace Congress
ISBN number: 978-1-925627-66-4
Please select category below:
Normal Paper
Student Paper
Young Engineer Paper
An industrial unsupervised Machine Learning model
combined with a signal processing approach to detect
failures in complex rotating assemblies
Michel Boussemart
1
, Mark Shariat
1
1
IoT Consultants, Lot Fourteen, Frome Road, Adelaide, 5000, SA, Australia
Abstract
To deploy Health Condition Monitoring (HCM) systems, which aim to realise early warning,
alarm and anomaly detection according to online data, there are numerous signal processing
methods to extract relevant information. Machine Learning techniques also appear in a
growing literature to avoid handcrafted analysis. However, they are challenging to train as
they require a large amount of labelled data representing healthy and non-healthy conditions.
Notably, applying purely data-oriented algorithms in an industrial context is challenging due
to a shortage of data from a machine running in an unhealthy state.
To address this challenge, an innovative framework is proposed, which is based on an
unsupervised machine learning framework combined with advanced signal processing
techniques, for detecting cyclo-stationary phenomena, well suited for rotating machines. This
framework is applied to publicly available datasets, showing very promising results on early
failure detection without training labelled data, i.e. only based on the indication of healthy
condition status.
Finally, the framework is applied to the HUMS2023 Data Challenge to detect as early as
possible, the planet gear crack before its propagation causes catastrophic consequences.
Keywords: Health Condition Monitoring (HCM), Unsupervised Deep Learning, Detection of
Cyclo-stationary signals.
Introduction
As planetary transmission systems become more and more widely used in complex
mechanical systems such as heavy trucks and helicopters, there is a significant research
interest in Prognostics and Health Management (PHM) and, in particular, in Health Condition
Monitoring (HCM) which aim to realize early warning, alarm and anomaly detection
according to online data. There are numerous methods to extract time or time-frequency-based
diagnostic features from input vibration signals, such as mean, RMS, Skewness, Kurtosis, and
spectral diagrams. Machine Learning techniques also appear in a growing literature to avoid
handcrafted analysis. However, they are challenging to train as they require a large amount of
labelled data representing healthy and non-healthy conditions. Notably, applying purely data-
oriented algorithms in an industrial context is challenging due to a shortage of data from a
machine running in an unhealthy state.
