Ship Engine Anomaly Detection Model

Business context

Dataset Url

I was provided with a real dataset to identify anomalous activity in a ship’s engine functionality (Devabrat, 2022). As I worked through this project, I kept in mind that, typically, anomalies make up a minority of the data points (i.e., about 1% to 5% of the data points were anomalies).

The dataset contained six critical features continuously monitored to evaluate the engine’s status as either ‘good’ or ‘bad.’ These features were:

• Engine rpm (revolutions per minute): I analyzed rpm levels, as high rpm indicated the engine was operating at a higher speed than designed for prolonged periods, potentially leading to overheating, excessive wear, and eventual failure. Conversely, a low rpm could signal a lack of power, fuel delivery issues, or internal mechanical problems.
• Lubrication oil pressure: Low lubrication oil pressure indicated insufficient lubrication, leading to increased friction, overheating, and engine damage. High lubrication oil pressure could signal a blockage in the oil delivery system, potentially causing seal or gasket failure.
• Fuel pressure: I monitored high fuel pressure, which could cause poor engine performance and incomplete combustion, suggesting fuel pump or filter issues. Low fuel pressure could result in excessive fuel consumption, poor emissions, or damage to fuel injectors.
• Coolant pressure: Low coolant pressure indicated potential leaks in the cooling system or coolant pump failure, which could lead to engine overheating. High coolant pressure signaled a possible blockage in the cooling system or a failing head gasket, both of which risked overheating.
• Lubrication oil temperature: High lubrication oil temperature suggested overheating, which could degrade the oil’s lubricating properties and lead to engine damage. Low lubrication oil temperature could indicate that the oil was not reaching its optimal operating temperature, leading to inadequate lubrication.
• Coolant temperature: High coolant temperature signaled overheating, possibly caused by a failed thermostat, coolant leak, or insufficient coolant flow. Low coolant temperature could indicate that the engine was not reaching its optimal operating temperature, reducing performance and efficiency.

Issues with engines could lead to malfunctions, safety hazards, and downtime, which could delay deliveries, impact a ship’s overall functionality, and harm the business through lost revenue. By predicting timely maintenance, I aimed to help the business increase profits by reducing downtime, mitigating safety risks for the crew, limiting fuel consumption, and improving customer satisfaction through timely deliveries.

Since I had not yet met with relevant business stakeholders (e.g., the ship’s engineers), I carefully considered the business context while detecting anomalies. I recognized that anomalies detected in the data might not always directly translate to operational issues without understanding the underlying business context. For example, while conducting exploratory data analysis and applying statistical techniques like the interquartile range (IQR), I might identify outliers indicating high revolutions per minute (RPM).

However, I understood that high RPM in isolation might not necessarily indicate a problem. For instance, it could signify the ship accelerating under appropriate conditions. To address this, I flagged anomalies only when specific combinations of features occurred simultaneously. For example, a high RPM coinciding with high coolant pressure and elevated lubrication oil temperature might indicate a potentially problematic scenario requiring attention. This approach ensured that I analyzed not just individual feature outliers but also interactions and combinations of features to accurately discern anomalous activity.

This project underscored the importance of understanding the intricacies of the ship’s operations and the interplay between various sensor readings. Without direct input from the ship’s engineers or domain experts, I approached anomaly detection with a keen awareness of the nuanced relationships between different features and their potential implications for operational integrity.

My task was to develop a robust anomaly detection system to protect the company’s shipping fleet by evaluating engine functionality. To achieve this, I explored the data, applied preprocessing and feature engineering, and detected possible anomalies.

Finally, I prepared a report illustrating my insights for the prospective stakeholders, showing how my solution could save the business money and build trust with its stakeholders. At this stage of the project, I focused on addressing the following key question:

• What insights could I gain from the data, and what recommendations could I make to the company based on these insights?
  For example:
  • Which features (and their thresholds) must be monitored closely for anomaly detection?
  • Which technique (statistical or ML) is the best for anomaly detection based on this dataset?
  • Which feature (univariate approach) or combination of features (multivariate approach) could predict maintenance?

Report on Findings