A Data-Driven Time-Series Fault Prediction Framework for Dynamically Evolving Large-Scale Data Streaming Systems

This paper presents a framework to tackle the problem of real-time fault classification for systems that can change their operating conditions dynamically over time. The observation about the system comes in a large-data streaming way. This framework is based on a nonparametric fuzzy-based classifier built to deal with large-scale, streaming, and dynamically evolving problems called Self-Organizing Fuzzy (SOF) classifier. The proposed approach is divided into two main stages: in the first stage, a feature extraction algorithm and a hypothesis test based on the Kolmogorov-Smirnov test are used to extract and select features from the observed time series. In the second stage, the SOF classifier processes the selected features to determine the existence of a fault pattern in the time-series data stream. The approach presented in this paper was tested in the real-world problem of real-time fault classification of railway switch machines. The obtained results show that the SOF classifier is more effective and has a lower computational cost than alternative approaches in the literature.