Unsupervised Autoencoder-Based Anomaly Detection for Storm Formation Using Onboard Time-Series Data  

Monday, April 27, 2026 at 2:00pm to 3:00pm

Advisor: Dr. Haiping Xu
Committee Members: Dr. Adnan El-Nasan and Dr. Joshua Carberry
 
Abstract: Storms remain among the most serious hazards in maritime environments, where rapidly changing weather conditions can threaten human life and disrupt vessel operations. Existing storm detection and forecasting methods often rely on satellite observations and large-scale numerical weather models, both of which require substantial computational resources and stable communication links that may be unavailable at sea. In addition, many machine learning approaches depend on large labeled datasets and predefined storm categories, limiting their ability to detect rare, rapidly developing, or previously unseen events. This thesis presents an unsupervised, sensor-based framework for early storm formation detection using multivariate meteorological time-series data collected from onboard instruments. The proposed approach uses directly measurable variables, including surface pressure, wind speed, sea surface temperature, near-surface air temperature, and dew point temperature, without relying on external data sources. A temporal autoencoder is trained exclusively on fair-weather observations to learn the normal temporal behavior of atmospheric conditions. During real-time operation, deviations from this learned baseline are quantified through reconstruction error, enabling the system to identify anomalous patterns associated with early storm development. To improve reliability in practical deployment, the framework incorporates persistence-based alerting and statistically derived anomaly thresholds based on fair-weather variability. The system is evaluated using both synthetic storm scenarios and real storm events derived from historical storm archives. Results demonstrate that the proposed method consistently detects storm formation with reasonable lead times and provides meaningful early warnings in practice, making it a practical and robust solution for enhancing maritime safety in remote or resource-limited environments.
 
All CIS graduate students are encouraged to attend. For further information please contact Dr. Haiping Xu at hxu@umassd.edu

Dion 311
Haiping Xu
5089998457