Performance Comparison of Machine Learning Models for Multi-Event Classification of Vibrational Patterns in Distributed Acoustic Sensing System

The phase-sensitive optical time domain reflectometer has extensive vibration detection applications in various domains. It can detect multiple vibrational event patterns due to its distributed sensing nature. Effectively classifying these events based on their features holds significant importance in distinguishing between different types of vibrations. This work focuses on showcasing the exceptional characteristics of various machine learning models for event classification, prioritizing computational efficiency and reduced processing time. The emphasis is also placed on the judicious selection of features to optimize classification accuracy. In this work, a publicly available dataset with six vibrational events was chosen and the classification of events was performed using three machine learning-based classifiers - Support Vector Machine, Random Forest, and K-Nearest Neighbor, and three machine-learning gradient boosting algorithms - Extreme Gradient Boosting, Categorical Boosting, and Light Gradient Boosting Machine. Some of the key hyper-parameters of the algorithms were also varied to understand the trade-off between the accuracy and computational complexity of the models. The Light Gradient Boosting Machine achieved the highest accuracy of 95% with fewer estimators and computation time, which was evaluated by the appropriate tuning of the hyperparameters.