Anomaly Detection in Unattended Machinery Plant Using an Bi-LSTM-DVAE Approach

Abstract This study proposes an anomaly detection framework based on a Bi-directional Long Short-Term Memory Disentangled Variational Autoencoder (Bi-LSTM-DVAE) to monitor and analyze operational data from a vessel. By leveraging the sequential modeling strength of Bi-LSTM networks and the probabilistic latent space of Disentangled Variational Autoencoders, this model captures complex temporal dependencies and representations of typical operational conditions. Trained on high-resolution time-series data, including metrics such as engine speed, fuel consumption, and lubricant oil pressure, the Bi-LSTM-DVAE model identifies anomalies by detecting reconstruction errors and deviations in latent space distribution. These anomalies can indicate emerging mechanical issues or shifts in operational patterns, allowing for early intervention and minimizing downtime. The framework demonstrates strong potential for enhancing anomaly detection and predictive maintenance in marine vessels, contributing to improved operational safety and reliability in the challenging conditions of maritime environments.