Analysis of time series data

Abstract

Nowadays emergence of the innovative approaches in data science and machine learning, enforces their use in modeling of the real world physical problems, even when they have been already explored and modeled with solid results. The evaluation of a partial discharge activity, as a phenomenon implying malfunction on an observed system, is one of such problems. The motivation of its reexamination and use of new data science approaches is motivated to increase the relevance of extracted knowledge which will be beneficial for the overall detection performance. The original data obtained by a patented metering device, deployed in the real environment, only underlines this need.

This thesis deals with an analysis and feature extraction from the time series data in order to design a robust fault detection mechanism. The robustness means the ability to correctly process an input data with various defects and interferences while focusing only on what is relevant and to gather as much valuable information about it as possible.

The entire work is a set of experimental models and analyses interconnecting a fundamental knowledge of the observed data with modern bio-inspired and soft-computing based machine learning algorithms and optimization approaches. The referential solution inspired by a state-of the art knowledge is designed with adjustable feature extraction process which parameters are further optimized making use of an swarm based optimization. Another models using evolutionary based feature synthesis, wavelet based signal decomposition or denoising driven by weighted singular values serve as the competitors in order to reveal other possibilities in studied problem.

The estimation of entropy, complexity and chaos in the data was supposed to increase the set of applicable features for the detection. The separability of several complexity indicators, like sample entropy, approximate entropy, 0-1 test for chaos and correlation dimension, was examined on data containing all kinds of measured malfunctions. Gathered results were accompanied with a discovery of a significant instability of one testing indicator, which has been found and reported for the first time.

The another author's proposals to represent the partial discharge pattern as a complex network are also novelties and they brought superior results in comparison with the state-of the art based classification models. They solid reasoning and simplicity offered multiple optimizations and evaluations which are documented in this work.