Data-Driven Model for Maintenance of Uninterrupted Power Systems
No Thumbnail Available
URL
Journal Title
Journal ISSN
Volume Title
Sähkötekniikan korkeakoulu |
Master's thesis
Authors
Date
2019-03-11
Department
Major/Subject
Control, Robotics and Autonomous Systems
Mcode
ELEC3025
Degree programme
AEE - Master’s Programme in Automation and Electrical Engineering (TS2013)
Language
en
Pages
62
Series
Abstract
A uninterruptible power system (UPS) is a device that can replace the continuous supply of power from the mains when the power is cut off. Its power comes from the battery pack. Eaton is a world leading provider of quality backup power UPS. Eaton UPSs are designed to deliver high-quality backup power from desktop PCs to large data centers. In order to keep all of the UPSs those are sold to customers work well, every sold UPS should be monitored to avoid the damages caused by failure events. In ’Connecting Marine Uninterruptible Power Systems to the Internet of Things [1], Esko Sorvato proposed the method to augment the predictive maintenance functionality in UPS through cloud computing. Cloud computing is an efficient method to do remote monitoring. The method is reliable, manageable, and economical. However, it is still suffering from high computing power, large storage requirements, large data transferring and complex failure predicting. This thesis provides a new method to select important features from a large-scale original datasets based on gradient boosting classifiers and develops an anomaly detection data-driven model for temporal datasets base on Temporal Convolutional Neural Network (TCN) method and long short-term Memory within the area of deeping learning. Firstly, in feature selecting, The approach to select the frequency is described in order to keep the original information as much as possible. Gradient boosting classifier is built to return the rank of important list. Secondly, TCN and RNNs with long short term memory (LSTM) units with the ability to predict the future behaviors if UPS is trained according to the normal time-series pattern. Gaussian Mixture Distribution (GMD) method is used to get the prediction errors and modeled to give anomaly scores. By comparing these two different kinds of data-driven models, TCN and anomaly detection are selected as the final anomaly detection model because of the better performance on temporal datasets on artificially generated failure datasets. With the experiments on different kinds of synthesis datasets, the results show that the size of datasets after feature selection is only 1/1000 of original one and TCNs are suitable for anomaly detection and time series modeling for the collected UPS datasets. Moreover, it is still worth to achieve LSTMs on real-world failure datasets.Description
Supervisor
zhou, quanThesis advisor
Sjoberg, AndersKeywords
data-driven model, anomaly detection, CNN, TCN, RNN, UPS