Simulating time-series data for improved deep neural network performance

Files

hdl_122754.pdf (10.79 MB)
  (Published version)

Date

2019

Authors

Yeomans, J.
Thwaites, S.
Robertson, W.S.P.
Booth, D.
Ng, B.
Thewlis, D.

Editors

Advisors

Journal Title

Journal ISSN

Volume Title

Type:

Journal article

Citation

IEEE Access, 2019; 7:131248-131255

Statement of Responsibility

Jordan Yeomans, Simon Thwaites, William S. P. Robertson, David Booth, Brian Ng, Dominic Thewlis

Conference Name

Abstract

Deep learning algorithms have shown remarkable performance in classification tasks, however, they typically perform poorly with small training datasets due to overfitting. Overfitting occurs for all data types, although for the purposes of this study we are interested in time-based signals. This study introduces a novel technique to simulate time series signals from a dataset of categorically labeled data which can be used to train a deep neural network. The objective is to improve the predictive accuracy of a deep neural network on a separate validation dataset. To demonstrate the simulation methodology and improvements to the model's performance, a small dataset of ground reaction forces was used with the goal of identifying a person based on the raw signal. Our results show that the simulation method presented improves validation accuracy and reduces model training time for each of the three signal types.

School/Discipline

Dissertation Note

Provenance

Description

Access Status

Rights

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see http://creativecommons.org/licenses/by/4.0/

License

Call number

Persistent link to this record