A Deep Support Vector Data Description Method for Anomaly Detection in Helicopters

##plugins.themes.bootstrap3.article.main##

##plugins.themes.bootstrap3.article.sidebar##

Published Jun 29, 2021
Chenyu Liu Konstantinos Gryllias

Abstract

Helicopters are high-value mechanical assets which has gained much attention from condition monitoring practitioners. Modern helicopter health management system leverages various sensors to collect in-flight signals. In order to trigger the alarm when an anomaly happens, signal processing methods are used to construct health indicators that require expert knowledge. On the other hand, classic features are always case-specific and may fail to discriminate anomalous in practical applications. Support Vector Data Description (SVDD) is a machine learning method used as a one-class classifier to serve anomaly detection tasks. It utilizes healthy samples to construct a hyper-sphere feature space as a detection threshold. In order to automate the anomaly detection pipeline, a deep SVDD model is proposed in this paper. A Convolution Neural Network (CNN) is used as the feature extractor, which provides smart features to an SVDD model. The SVDD model uses a soft-boundary hyper-sphere for decision-making. The optimization of the CNN and the SVDD is connected, which makes it an end-to-end process. The methodology is applied, tested and evaluated on a helicopter vibration dataset, which has been provided by Airbus SAS in the frames of AI Gym challenge. The experimental results reveal that the F1 score of the proposed Deep SVDD can reach 94%, showing its compelling efficacy for anomaly detection.

How to Cite

Liu, C., & Gryllias, K. . (2021). A Deep Support Vector Data Description Method for Anomaly Detection in Helicopters. PHM Society European Conference, 6(1), 9. https://doi.org/10.36001/phme.2021.v6i1.2957
Abstract 1123 | PDF Downloads 1630

##plugins.themes.bootstrap3.article.details##

Keywords

Deep learning, Anomaly detection, Support Vector Data Description, Condition monitoring, Deep SVDD

Section
Technical Papers