Comparative Analysis of Fixed-Length and Dynamic Segmentation for Feature Extraction from Non-Stationary Spatial Data

This research presents a comparative analysis of non-stationary spatial data segmentation techniques such as fixed-length and dynamic segmentation based feature extraction efficiency. The study utilizes 5 miles of railway track geometry data, a non-stationary spatial dataset, to assess the effectiveness of both segmentation approaches. The profile (vertical alignment) of the track geometry is used for this purpose. For fixed-length segmentation, the track data is divided into segments of 264 feet (1/20th of a mile), resulting in about 102 segments. Dynamic segmentation is performed using an "l2" model-based change point detection algorithm, which adapts to natural variations in the signal. Key features such as standard deviation, kurtosis, and energy are extracted from both segmentation methods. Performance is evaluated based on multiple criteria, including the discriminative power of the features for classifying track safety and ride-quality conditions using statistical tests such as the f-test and Fisher score, consistency or signal quality across segments, measured using the variance of the signal-to-noise ratio (SNR), computational efficiency in terms of run-time and memory usage. Results indicate that, features from fixed-length segments have demonstrated better discriminative power between safety and ride quality classes, with higher Fisher scores and f-values showing strong statistical significance (p < 0.05). Additionally, fixed-length segmentation has shown a better performance with lower run-time and stable signal power across segments.