Time-Series Anomaly Detection for Tactical Vehicle Platforms

Modern combat vehicles like the M1 Abrams and M109 Paladin generate massive volumes of high-frequency, multi-variant sensor telemetry. Within these complex datasets, early indicators of subsystem degradation often manifest as transient, low-probability anomalies, such as a localized voltage variance sustained for an eight-second interval, which are computationally infeasible to isolate manually. This project evaluates an unsupervised machine learning framework designed for sparse anomaly detection in unlabeled time-series data. Specifically, this study analyzes classical, low-compute algorithms (such as Isolation Forests) alongside deep generative models (such as Autoencoders) to establish normal operational baselines. By evaluating how well these models detect these anomalies, the goal is to identify early component wear before physical fault codes are triggered. Ultimately, this framework provides an analytical decision-support tool to optimize maintenance scheduling, reduce unscheduled downtime, and maximize combat vehicle availability.