Comparison of Dimensional Reduction Methods for Predictive Analysis of Railway System Data

Predictive analysis uses meaningful correlations in subsets of datasets for identifying patterns to predict future behavior. Railway system operators acquire real-time data through Internet of Things (loT) sensors used in the protection and monitoring systems. In this work different fault scenarios are analysed through dimensionality reduction methods Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA) and Autoencoder (AE) with K-means clustering. The methods are evaluated on operative converter station dataset from the Norwegian railway system through performance metrics to relatively compare the clustering quality. The results indicate that the AE ‘ s performance is particularly good, achieving a Silhouette value of 0.625 for 3 clusters. Additionally, AE tends to maintain a lower Davies-Bouldin value. It is observed shows that PCA has its strengths, particularly in reducing dimensionality and still capture the most variance in the data, AE not only reduces the dimensionality but also learns a representation of the data that is better at capturing complex non-linear relationships, which can be more beneficial for the clustering task. Hence, based on the results, AE would be preferred over PCA and LDA if the goal is to obtain distinct and well-separated clusters.