This project's objectives were to develop an anomaly detection system to protect a company’s shipping fleet by evaluating engine functionality, ensuring early identification & malfucntion prevention.

The dataset used in this project consists of 19535 rows and 6 columns, earch representing different engines and their vital metrics including - Engine RPM, Lubricant oil pressure, Lubricant oil temperature, Fuel pressure, Coolant pressure & Coolant temperature.

The dataset can be found here: dataset

• Performed inital data exploration and preprocessing.
• Applied the interquartile range (IQR) method to identify outliers for each feature. Extracted the final sample of outlier where only two or more of the features fell under an outlier category.
• Instantiated and ran the OneClassSVM pipeline and experimented with parameter values to optimise the model's performance.
• Applied the Isolation forest algorithm and tested different combinations of parameter settings to improve the model's outlier predictions.
• Performed dimensionality reduction on the data using Principle component analysis (PCA) to enable simpler and more intuitive data visualisation.
• Compared the outputs of the two ML methods.

• Python, Numpy, Pandas, SciPy Stats
• MatPlotLib, Seaborn
• SciKitLearn (OneClassSVM, Isolation Forest & PCA)

• Both optimised machine learning methods were able to indentify the expected range of anomalies (1-5%).
• While the IQR method, when combined with a rule, produced results within the expected range, the multivariate machine learning methods (SVM and Isolation Forest) were more comprehensive in identifying anomalies by considering the relationships between features.
• The identified anomalies can serve as a basis for further investigation by qualified engineers and service personnel to determine the root cause and necessary maintenance.