A Dataset for Evaluating Online Anomaly Detection Approaches for Discrete Multivariate Time Series

The PATH dataset, introduced for evaluating online unsupervised anomaly detection methods in multivariate time series, simulates a realistic automotive powertrain, offering diverse, extensive, and non-trivial data. Available in contaminated and clean versions, it supports unsupervised and semi-supervised settings, time series generation, and forecasting. The dataset, along with source code, is accessible online. It features 16 logged signals from a full electric vehicle's subsystems, chosen based on domain knowledge to represent powertrain behavior. The study focuses on creating a diverse dataset with realistic drive cycles to assess the effectiveness of classical methods against deep learning in anomaly detection. The PATH dataset includes signals like motor speed, torque, battery state of charge, and pedal inputs, and is designed to test anomaly detection methods in the presence of contamination.