Clustering time-series data

Abstract

This thesis investigates the application of machine learning techniques to classify and cluster time-series data from motor vibrations, focusing on enhancing condition-based monitoring (CBM) systems for industrial automation.

The primary aims are to assess the data generation capabilities of a custom-built test rig and to evaluate the effectiveness of various classification and clustering methods on this data.

The study explores several techniques, including Dynamic Time Warping (DTW), Continuous Wavelet Transform (CWT), Long Short-Term Memory networks (LSTM), and convolutional neural networks such as ResNet18 and VGG19, paired with clustering algorithms like K-Means and K-Medoids.

The findings confirm that the test rig is effective for generating reliable data. While LSTM models excel in managing complex datasets with minimal variations, traditional classification methods like ResNet18 combined with CWT are highly accurate but struggle with subtler datasets. Clustering methods show promise but vary in effectiveness depending on the distinctiveness of the data. This research underscores the potential of advanced data processing techniques to improve the reliability and efficiency of CBM systems in industrial settings.