Highly Accurate Anomaly Detection for the Combustion System of a Marine Engine Using Multivariate Transformers

This paper proposes a novel anomaly detection system to identify combustion process irregularities, in real-time, within individual cylinders of marine engines. The system monitors the instantaneous speed of the crankshaft to detect deviations that might indicate abnormal power output from specific cylinders. This information is then communicated to the engine control system, enabling timely corrective actions. Due to cost, privacy, and availability constraints in acquiring real-world training data, we address this by generating syn-thetic crankshaft speed data using a vessel engine simulator. We further introduce real-world anomaly scenarios, via fault injection, simulating a single cylinder producing higher-than-normal RPM. We implement a transformer-based encoder model specifically designed for multivariate time series data to detect these anomalies. The model leverages self-attention mechanisms and positional encodings to capture temporal dependencies and interactions between variables within the crankshaft speed data. Our proposal achieves a high accuracy of over 96% in anomaly detection.