The signs are there, now predict the future! Predicting System Failure and Reliability

Drenth, R.

The signs are there, now predict the future! Predicting System Failure and Reliability

Files

Drenth, RJ.,_MA_Thesis_2014.pdf (1.45 MB)

Authors

Drenth, R.

Issue Date

2014-08-29

Language

en

URI

http://theses.ubn.ru.nl/handle/123456789/223

Abstract

All kinds of machines and electronics are finding their way into our everyday lives and we increasingly rely on their proper functioning. Especially when such machines are key to business operations, their reliability is vital. For parking garages this is also the case, where a non-functional machine can lead to both a short and long term loss of income. In order to guarantee that machines stay operational, a generic predictive setup was developed. The basis of this setup was formed by sequences of recorded events, which describe what happens to a device, including regular occurrences as well as system failures. Such a sequence of events are also known as time-series and our setup is aimed to predict future time-steps of these series, which describe whether each possible event will occur in the future. Such predictions can aid the process of deciding whether proactive maintenance should be performed or not. For our setup, we tested the Multilayer Perceptron and Conditional Restricted Boltzmann Machine algorithms, as well a Recursive and Single Step paradigm for time-series prediction, which determine how multiple future time-steps are predicted. We furthermore investigated the influence of different data pre-processing methods on the setup's performance, such as the addition of extra features to the dataset, which are generated using Principal Component Analysis, and the application of an outlier treatment method, which enforces a maximum value during the normalisation process. The two mentioned data pre-processing methods were proven to be ineffective, while we demonstrated that a setup utilising the Conditional Restricted Boltzmann Algorithm, in combination with the Single Step prediction paradigm, was able to create generalised prediction models for devices of different brands. Such a setup achieved similar average accuracy scores of almost 75% for the events within predicted time-steps of the time-series, regardless of a device's type or brand, which is sufficient for its intended real-world application.