PEER REVIEW
20
th
Australian International Aerospace Congress, 27-28 November 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
Embedding signal processing knowledge in neural
networks – An application to gear diagnostics
P. Borghesani
1
, N. Herwig
1
, W. Wang
2
and J. Antoni
3
1
School of Mechanical & Manufacturing Engineering, UNSW Sydney, Australia
2
Aerospace Division, Defence Science and Technology Group, Australia
3
Univ Lyon, INSA Lyon, LVA, EA677, 69621 Villeurbanne, France
Abstract
Knowledge-based signal processing methods for detecting faults using vibration data analysis
have been researched and applied in practice for a long time. In contrast to deep machine
learning methods, analytical methods do not require a data basis to determine their output.
However, this often requires expert intervention to adapt them to new applications. To automate
this process, deep machine learning (ML) methods in condition monitoring have been explored
in recent years. These methods require a large amount of data to perform accurate diagnostics.
However, in many practical applications, fault data is scarce, which limits the field of
application of deep ML methods. For this reason, physics-informed neural networks (PINN) are
attracting increasing attention. In PINNs, neural networks optimise parameters in a range
restricted by analytical methods. On the one hand, this ensures that the results are physically
meaningful, and considerably less data is required than with pure neural network approaches.
This paper presents a PINN for condition monitoring of vibrational data. Based on a simple
structure consisting of two convolutional layers and two dense layers, the network edits the
spectrum in an interpretable way from a knowledge-based signal processing point of view and
calculates a fault index that has a strong relationship to the average log-ratio (ALR). The
generalization capability of signal processing steps is automatically adapted to specific data
through implementation by an NN. The presented PINN methodology is benchmarked against
a black-box convolutional neural network (CNN) and a long-short-term memory (LSTM)
network. We show that PINN outperforms classical pure ML methods when applied to a small
data set (the UNSW TMCM’s gear-crack dataset). Furthermore, it is shown that the reason for
this result lies in the fact that black-box approaches, in contrast to the presented PINN, are not
able to identify the physically correct fault features for small datasets.
Keywords: condition-based maintenance, fault detection, neural network, rotating machinery,
signal processing
Introduction
Neural networks (NN) have been proven powerful tools in many fields. Success stories have
prompted a prolific new direction for machine condition monitoring (MCM) research.
However, significant class imbalance (few fault data) significantly complicates the application
of NNs in MCM [1]. Almost the entirety of MCM-NN research relying on minimal laboratory
sets, where overfitting is very likely, and the capability of the NN to generalise to other data is
