Hierarchical level fault detection and diagnosis of ship engine systems

As smart ships are developed, research into engine management through remote monitoring and support from data collected at onshore control centers is being conducted. However, ship engine data collected from various sensors have high dimensions, large measurement errors, few labeled data, and insufficient amounts and quality of data, making it difficult to determine the condition of engines with complex failure modes under various operating environments. This study proposes a hierarchical level fault detection and diagnosis (HL-FDD) method that combines domain knowledge of ship engines and advanced data analysis techniques. The developed method extracts key features of reduced dimensions from the original variables (sensors) through an optimal hierarchical clustering and dimension-reduction model, allowing a hierarchy divided into the top (the entire system combining all features), middle (subsystems and feature sets), and bottom (components and sensors) levels. The reduced key features are used to generate robust regression models and dynamic thresholds (prediction intervals) according to the engine load, and dynamic thresholds determine whether the engine's condition is normal or abnormal. The dynamic thresholds are able to automatically label abnormal conditions of the engine. Once anomalies are detected at the top level, the proposed method can sequentially search for data on features belonging to the middle and bottom hierarchies for detailed fault diagnosis to determine which engine subsystem or component(s) caused the engine fault(s). Actual data collected by ship operators verify the proposed method's efficiency, reliability, and accuracy.