Back to Projects
Anomaly Detection / Operations

Maritime Engine Anomaly Detection System

An intelligent monitoring system that detects subtle degradation patterns in ship engine operational data before conventional rule-based alerts would fire, reducing the risk of costly unplanned downtime.

Type
Anomaly Detection
Domain
Maritime Operations
Methods
Isolation Forest, PCA
Status
Completed
ANOMALY DETECTION
ISOLATION FOREST

The Challenge

Maritime engine systems generate continuous streams of operational data across dozens of sensors. Conventional monitoring relies on static thresholds, triggering alerts only when individual readings exceed pre-set limits.

Threshold-based monitoring misses gradual degradation. A slow drift in the relationship between sensor readings can indicate developing problems long before any single metric breaches a hard limit. By the time a threshold alert fires, the damage is often significant.

Approach

01
Multivariate Data Exploration
Explored the sensor dataset to understand normal operational patterns, inter-variable correlations, and natural variation ranges under healthy operating conditions.
02
Feature Engineering
Engineered features capturing temporal dynamics and cross-sensor relationships. Applied normalisation and PCA to create a compact representation preserving anomalous variance.
03
Anomaly Detection Modelling
Applied Isolation Forest as the primary detection method, complemented with statistical methods for interpretable cross-validation of findings.
04
Analysis and Reporting
Produced a detailed analytical report characterising detected anomalies, mapping them to potential operational causes, and recommending monitoring improvements.

Results

977
Anomalous episodes detected across the operational dataset
19,535
Sensor readings analysed across 6 operational channels
6 → 2
PCA dimensionality reduction while preserving anomalous variance

Analysed 19,535 sensor readings across six operational channels, applying PCA to reduce the feature space from 6 dimensions to 2 while preserving the variance most indicative of anomalous behaviour. Isolation Forest then identified 977 anomalous episodes - subtle multi-sensor degradation patterns that single-threshold monitoring systems miss entirely.

By analysing cross-sensor relationships rather than individual channel limits, the system detects anomalous behaviour during the gradual drift phase - before individual readings breach hard limits. In operational environments where unplanned engine failure carries significant cost and safety consequences, the difference between threshold-based alerts and pattern-based early detection is the difference between reactive repair and planned intervention.

The system was designed for interpretability: detected anomalies were mapped to potential operational causes with supporting analytical evidence, enabling engineering teams to investigate and act rather than simply receive opaque alerts.

Technology Stack

Python Isolation Forest PCA Scikit-learn Pandas Matplotlib Seaborn
Interested in this work or something similar?