Session: 02-12-01 Reliability Based Maintenance and Inspection Planning

Submission Number: 157386

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

This study presents a robust anomaly detection framework utilizing a Long Short-Term Memory Variational Autoencoder (LSTM-VAE) to monitor and analyze operational data from a marine vessel. Maritime operations often involve highly dynamic environments, making the timely identification of anomalies critical for ensuring operational safety and efficiency. The LSTM-VAE combines the sequential modeling capabilities of Long Short-Term Memory (LSTM) networks with the probabilistic latent space of Variational Autoencoders (VAE) to capture complex temporal dependencies and latent representations of normal operational behavior.

The model is trained on high-resolution time-series data, including engine speed, fuel consumption, and coolant temperature, capturing critical operational parameters. Anomalies are detected by analyzing reconstruction errors and deviations in the latent space, effectively identifying both short-term and long-term deviations from normal patterns. These anomalies provide early warnings of potential mechanical failures or inefficiencies, enabling proactive interventions and minimizing downtime.

This framework demonstrates resilience to noisy data, adaptability to diverse operating conditions, and the ability to capture subtle variations that traditional methods might miss. By leveraging insights from the latent space, it supports predictive maintenance and decision-making processes, offering significant operational advantages.

The results highlight the LSTM-VAE model’s potential to advance anomaly detection in maritime systems, enhancing reliability, safety, and efficiency in challenging operational conditions. This work underscores the importance of integrating advanced machine learning approaches to address the unique challenges faced in maritime environments, paving the way for future advancements in anomaly detection and predictive maintenance.

Presenting Author: Ahmad Bahootoroody Aalto University

Presenting Author Biography: Dr. Ahmad BahooToroody is an Academy of Finland Research Fellow in the Marine and Arctic Technology Group of Aalto University. With a Ph.D. in Reliability Engineering from the University of Florence (Italy), he was invited to liaise with researchers at TU Delft (The Netherlands), Norwegian University of Science and Technology (Norway) and University of Strathclyde (United Kingdom). His PhD research focused on risk-based asset integrity modeling of the automotive process, and he completed both the comprehensive exam and oral defense with distinction. Ahmad’s research interests stand at uncertainty quantification, trustworthiness of engineering operations, condition-based maintenance optimization, and application of probabilistic methods and machine learning tools in dynamic modeling.