Anomaly Detection for Marine Engines: Unsupervised Learning Methods Considering Sensor Noise and Time-Series Modeling

Abstract Marine engines are core components of ship propulsion systems, and their reliability is critical to the long-term stable operation of ships. However, due to their longterm operation in complex marine environments-exposed to high-salt fog and complex operating conditions-undetected anomalies can lead to fault escalation, posing significant risks to personal safety and property. Accurate anomaly detection is therefore crucial for ensuring their safe and stable operation. Practical data collected from marine engines are typically characterized by long-term time-series, and due to their complex operational environments, sensors inevitably introduce inherent noise. To address the anomaly detection needs of marine engines, this paper proposes an unsupervised anomaly detection framework, ITransformerVAE-HMM, which integrates latent variable modeling and time-series analysis. This method enables effective identification of anomaly states across multiple operating conditions without relying on labeled data. By incorporating a denoising autoencoder (DAE) mechanism into the model structure, it effectively reduces noise interference in raw measurement signals. Furthermore, combining the Transformer module with latent variable modeling enables the model to learn more discriminative features in the latent space. Finally, the Hidden Markov Model (HMM) is employed to model the state transition patterns of marine engines, thereby enabling the identification and prediction of anomaly states. Using the operating data of marine natural gas engines, we conducted a comparison and ablation study on the proposed method. The research results show that this method performs exceptionally well, with an accuracy rate of 99.69% and a false alarm rate of only 0.375%, outperforming the baseline method.