Ship Engine Anomaly Detection
A statistical rule and two unsupervised models compared for flagging abnormal ship-engine sensor readings, as candidates for engineering review.
The Challenge
An IQR rule flagged 422 of 19,535 ship-engine readings (2.16%) where at least two sensors were out of range, giving engineers a short list of readings to review.
Engine faults are expensive at sea, and sensor data arrives faster than engineers can review it. With no labelled faults, the question was how to flag a manageable set of unusual readings, and how statistical and machine-learning methods compare.
Approach
Results
| Method | Setting | Readings flagged |
|---|---|---|
| IQR rule (2+ sensors out of range) | - | 422 (2.16%) |
| One-Class SVM | nu = 0.05 | 980 (5.02%) |
| Isolation Forest | contamination = auto | 2,892 (14.8%) |
| Isolation Forest | contamination = 0.05 | 977 (5.0%) |
The IQR rule is the only result not fixed in advance. For One-Class SVM and Isolation Forest, the nu and contamination settings decide roughly what share of readings is flagged, so their 5% figures reflect the chosen setting rather than a discovery. Readings flagged by several methods are the strongest candidates for review.
Limitations: no labelled faults, so none of the methods can be scored for accuracy; readings are treated independently, without time order.
How it was tested
- Data reviewed before modelling: 19,535 readings from six sensors, no missing or duplicate rows.
- A statistical rule compared with two unsupervised models at several settings, and the flagged readings plotted with PCA.
- For One-Class SVM and Isolation Forest, the settings decide roughly what share is flagged, so their 5% figures reflect the chosen setting rather than a discovery. Only the IQR result is not fixed in advance.
- No labelled faults, so none of the methods can be scored for accuracy.
- Readings are treated independently, without time order.