Intelligent time-series forecasting and event prediction for Predictive Maintenance in IT systems

Pedro Ribeiro Santiago Moreira

Key information

Authors:

Pedro Ribeiro Santiago Moreira (Pedro Ribeiro Santiago Moreira)

Supervisors:

Ricardo Carvalho; João Paulo Baptista de Carvalho (João Paulo Baptista de Carvalho)

Published in

05/24/2021

Abstract

Nowadays, most Information Technology systems already perform Condition-based Maintenance, which provides an overview of a system's condition in real-time and stores its behaviour (in the form of time-series). A Predictive Maintenance approach can use this historical data to apply planning corrective maintenance based on predictions about a system's evolution. This work intends to provide useful research in the scope of Predictive Maintenance through the study of the best approaches and algorithms to perform time-series forecasting and event prediction within the Information Technology domain. The state-of-the-art intelligent methods for time-series modelling were studied, as well as the most promising methods developed for other domains with existent literature ahead of time-series (like Natural language Processing). The main focus was on Machine Learning techniques - from the primary Feed-forward Neural Networks all the way to the more recent and complex Transformers. For the time-series forecasting, none of the experimented models performed satisfactorily. However, it was notable that with the increase of complexity and size of the model's architectures, they learned to output "dummy" forecasts like a naive approach or a constant value, which although not useful, minimise the evaluation metrics. For the event prediction problems, a preprocessing step to detect oscillations in the input datasets significantly boosted the algorithms' performances. Furthermore, the results obtained were not ideal but satisfactory enough to be useful, and the model that showed the best results was the Feed-forward Neural Network. Finally, it is possible to adjust the predictions' sensitivity with the tuning of a data preprocessing factor.