Authors:
José Cação
1
;
2
;
Mário Antunes
3
;
4
;
José Santos
1
;
2
and
Miguel Monteiro
5
Affiliations:
1
TEMA, Centro de Tecnologia Mecânica e Automação, Departamento de Engenharia Mecânica, Univerisdade de Aveiro, 3810-193 Aveiro, Portugal
;
2
LASI, Laboratório Associado de Sistemas Inteligentes, Guimarães, Portugal
;
3
DETI, Departamento de Eletrónica, Telecomunicações e Informática, Universidade de Aveiro, Campus de Santiago, 3810-193 Aveiro, Portugal
;
4
IT, Instituto de Telecomunicações, Aveiro, 3810-193 Aveiro, Portugal
;
5
Bosch Termotecnologia S.A., 3800-627 Cacia, Portugal
Keyword(s):
Dimensionality Reduction, IIoT, Feature Extraction, Feature Selection.
Abstract:
The industrial landscape is undergoing a significant transformation marked by the integration of technology and manufacturing processes, giving rise to the concept of the Industrial Internet of Things (IIoT). IIoT is characterized by the convergence of manufacturing processes, smart IoT devices, and Machine Learning (ML) algorithms, enabling continuous monitoring and optimisation of industrial operations. However, this evolution translates into a substantial increase in the number of interconnected devices and the amount of generated data. Consequently, with ML algorithms facing an exponentially growing volume of data, their performance may decline, and processing times may significantly increase. Dimensionality reduction (DR) techniques emerge as a viable and promising solution, promoting dataset feature reduction and the elimination of irrelevant information. This paper presents a comparative study of various DR techniques applied to a real-world industrial use case, focusing on th
eir impact on the performance and processing times of multiple classification ML techniques. The findings demonstrate the feasibility of applying DR: for a Logistic Regression classifier, minor 4% performance decreases were obtained while achieving remarkable improvements, over 300%, in the processing time of the classifier for multiple DR techniques.
(More)