PEER REVIEW
20
th
Australian International Aerospace Congress, 27-28 February 2023, Melbourne
Please select category below:
Normal Paper
Student Paper
Young Engineer Paper
Automated feature selection for multi-channel anomaly
detection
Leonard Whitehead
1
, John Taylor
1,2,3
, Wenyi Wang
1
and Biswajit Bala
1
1
Defence Science and Technology Group, Department of Defence, Melbourne Australia
2
CSIRO Data61, Canberra, Australian Capital Territory, Australia
3
College of Engineering and Computer Science, The Australian National University,
Canberra, Australia
Leonard.whitehead@defence.gov.au, John.Taylor@data61.csiro.au
Abstract
Identifying and choosing the ideal feature inputs to a machine learning algorithm is a slow and
complex task. Methodologies such as univariate techniques, dimensionality reduction
techniques, stepwise selection techniques and expert knowledge are often used in evaluating
the input features. These techniques, although often effective and powerful, are not considered
to be automated feature selection techniques, because they require a thorough assessment and
interpretation by the user. This paper introduces a search algorithm using unsupervised
learning, which is built upon our previously developed methodology for applying machine
learning to detecting faults in fielded machinery. The search algorithm explores the input
feature space and selects the input feature(s) that are highly sensitive to anomalies within the
given dataset. The algorithm uses a heat map that shows the sensitivity of the feature inputs to
anomaly detection, which can be further developed into a fingerprint analysis method to
isolate faults within mechanical systems. Currently, this algorithm is being tested using real
world aerospace data for anomaly detection. Preliminary results are presented in this paper and
the final analysis results will be reported in a future paper.
Keywords: Anomaly detection, automated feature selection, condition monitoring, fault
fingerprinting, machine learning.
Introduction
Input features are known to have a direct effect on the performance of machine learning models.
Feature selection (or variable elimination) is the process of selecting variables which efficiently
describe the input data, reduce the effects of noise and provide good prediction results [1]. The
automated feature selection for multi-channel anomaly detection is a wrapper-based feature
selection algorithm, which uses an exhaustive search strategy (or grid search) with a two-sample
Kolmogorov-Smirnov (KS) test for filtering purposes. The induction algorithm (i.e. the neural
network) has been built on our previous research, which used the Levenberg-Marquardt (LM)
optimizer [2]. The algorithm leads to a fingerprinting analysis via a heatmap which identifies
