Detail publikace

Data pre-processing effect on classification accuracy of convolutional neural networks for train type identification

KRČ, R. PODROUŽEK, J. VUKUŠIČ, I. PLÁŠEK, O.

Originální název

Data pre-processing effect on classification accuracy of convolutional neural networks for train type identification

Anglický název

Data pre-processing effect on classification accuracy of convolutional neural networks for train type identification

Jazyk

en

Originální abstrakt

Accelerometer data collected by in-situ measurements near the common crossing from two locations in the Czech Republic were used for training and validation of machine learning models. Several architectures of convolutional neural networks (CNN), successfully applied in the previous research for electrical grid analysis, were evaluated in this paper for the problem of locomotive type identification with regards to the number of parameters and the size of the available dataset, which was limited in this case. Therefore, time-series pre-processing techniques aiming to improve classification accuracy by removing noise were incorporated, including Butterworth low-pass and high-pass filters as well as Wavelet threshold filter. Results for raw and filtered data are presented as mean confusion matrices to evaluate the statistical significance of the adopted methods.

Anglický abstrakt

Accelerometer data collected by in-situ measurements near the common crossing from two locations in the Czech Republic were used for training and validation of machine learning models. Several architectures of convolutional neural networks (CNN), successfully applied in the previous research for electrical grid analysis, were evaluated in this paper for the problem of locomotive type identification with regards to the number of parameters and the size of the available dataset, which was limited in this case. Therefore, time-series pre-processing techniques aiming to improve classification accuracy by removing noise were incorporated, including Butterworth low-pass and high-pass filters as well as Wavelet threshold filter. Results for raw and filtered data are presented as mean confusion matrices to evaluate the statistical significance of the adopted methods.

Dokumenty

BibTex 


@inproceedings{BUT171770,
  author="Rostislav {Krč} and Jan {Podroužek} and Ivan {Vukušič} and Otto {Plášek}",
  title="Data pre-processing effect on classification accuracy of convolutional neural networks for train type identification",
  annote="Accelerometer data collected by in-situ measurements near the common crossing from two locations in the Czech Republic were used for training and validation of machine learning models. Several architectures of convolutional neural networks (CNN), successfully applied in the previous research for electrical grid analysis, were evaluated in this paper for the problem of locomotive type identification with regards to the number of parameters and the size of the available dataset, which was limited in this case. Therefore, time-series pre-processing techniques aiming to improve classification accuracy by removing noise were incorporated, including Butterworth low-pass and high-pass filters as well as Wavelet threshold filter. Results for raw and filtered data are presented as mean confusion matrices to evaluate the statistical significance of the adopted methods.",
  booktitle="Computational Science and AI in Industry (CSAI 2021)",
  chapter="171770",
  howpublished="online",
  year="2021",
  month="june",
  pages="1--1"
}