Sensor degradation detection using visual timeseries and deep convolutional neural networks

Wallace, Christopher and McArthur, Stephen; (2021) Sensor degradation detection using visual timeseries and deep convolutional neural networks. In: Proceedings of 12th Nuclear Plant Instrumentation, Control and Human-Machine Interface Technologies (NPIC & HMIT 2021). American Nuclear Society, [S.I.], pp. 304-312. (https://doi.org/10.13182/T124-34395)

[thumbnail of Wallace-McArthur-ANS-2021-Sensor-degradation-detection-using-visual-timeseries-and-deep]
Preview
 Text. Filename: Wallace_McArthur_ANS_2021_Sensor_degradation_detection_using_visual_timeseries_and_deep.pdf
Accepted Author Manuscript
Download (1MB)| Preview

Abstract

Effective condition monitoring and prognostic analyses depend on accurate sensor data in order to estimate the health of an asset. Given the nature of many items of legacy equipment in the nuclear industry, this is often not the case, with degradation occurring due to the age and operation of sensors, communication routes as well as the asset itself. Detection of anomalous signals which do not reflect the behaviour of the asset is therefore essential, with the primary objective to repair or replace such components. This paper introduces a method of detection of sensor degradation using timeseries which have been converted to images, in order to leverage the powerful feature detection capabilities of modern deep convolutional neural networks. By converting 1-D time-series to 2-D representations via a Grammian Angular Field and using a small number of training examples, it is possible to train such a network to automatically identify features associated with faults. A case study is presented for a set of sensor types, demonstrating the capability of the model to generalise to previously unseen data from sensors of the similar type and identify faults at greater than 85% accuracy. The results demonstrate the benefits that can be derived from an unsupervised feature detection process for this type of problem and highlights the transferability of models trained on one sensor type and applied to previously unseen similar sensor types.

ORCID iDs

Wallace, Christopher and McArthur, Stephen ORCID logoORCID: https://orcid.org/0000-0003-1312-8874;
CORE (COnnecting REpositories)