OODCN: Out-of-distribution detection in capsule networks for fault identification

Imene Mitiche, Alireza Salimy, Falk Werner, Philip Boreham, Alan Nesbitt, Gordon Morison

Research output: Chapter in Book/Report/Conference proceedingConference contribution

31 Downloads (Pure)

Abstract

In order to aid survey engineers identify Partial Discharge (PD) types during their asset diagnostics, we develop an image-based system for PD signal classification and Out-Of Distribution (OOD) rejection. First, the PD signal is converted to a Phase-Resolved PD (PRPD) image. Then, the image is passed to the system which exploits a Capsule Network in an auto-encoder framework, where the encoder output is used for PD classification and the decoder output is used in the OOD decision. The latter is the main contribution of this work which combines the decoder part with a reconstruction metric evaluating the difference between the original and reconstructed image. A threshold for OOD decision is introduced based on the distribution of reconstruction values from the training data.Most importantly, the OOD data is not exposed to the model during training. Results demonstrate high performance in both PD types classification and OOD detection tasks using synthetic and real data.
Original languageEnglish
Title of host publication2021 29th European Signal Processing Conference (EUSIPCO)
PublisherIEEE
Pages1686-1690
Number of pages5
ISBN (Electronic)9789082797060
ISBN (Print)9781665409001
DOIs
Publication statusPublished - 8 Dec 2021
Event29th European Signal Processing Conference - Online
Duration: 23 Aug 202127 Aug 2021
https://eusipco2021.org/ (Link to conference website)

Publication series

Name
ISSN (Print)2219-5491
ISSN (Electronic)2076-1465

Conference

Conference29th European Signal Processing Conference
Abbreviated titleEUSIPCO 2021
Period23/08/2127/08/21
Internet address

Keywords

  • partial discharges
  • training
  • measurement
  • rotating machines
  • training data
  • signal processing
  • data models

Fingerprint

Dive into the research topics of 'OODCN: Out-of-distribution detection in capsule networks for fault identification'. Together they form a unique fingerprint.

Cite this