Bow slamming detection and classification by Machine Learning approach

Identification of slamming events during ship navigation is typically based on the fulfillment of multiple conditions evaluated on sensor data. Alternatively, Machine Learning (ML) algorithms can be trained to capture slamming events by analyzing features from measurement signals. The feasibility of using supervised ML techniques is here presented by processing data from a segmented-model of a fast ferry tested in the towing-tank. The extensive test conditions provided a suitable dataset for ML training and comparing the accuracy of results with a physics-based identification model. The slamming identification problem is addressed at two levels: simply counting of slams, and classification of slams into three different groups. A challenging aspect is related to imbalanced data, due to minority of slamming classes, calling for meaningful evaluation metrics. The considered ML models are eXtreme Gradient Boosting (XGBoost), Supporting Vector Machine and Decision Tree for the detection problem, along with meta-models to deal with classification. A maximum F1-score of 72.4% is reached in slamming detection, while the best weighted F1-score for classification is 84.3%, both obtained with XGBoost. Reducing the number of analyzed features, as in the case of sensor failure, still provides good performance metrics, demonstrating the versatility of the ML approach.