Machine Learning-Based Discrimination of Pre- and Post-Cleaning Marine Turbocharger Recordings Using Rotational-Speed Signals

Monitoring marine turbocharger condition under real-ship operation is complicated by operating-condition variability and limited labelled maintenance data. This study assesses whether mean-centred angle-domain rotational-speed fluctuations can discriminate between six recordings acquired before and after one turbocharger cleaning event. The primary evaluation used exhaustive paired complete-record holdout, where one pre-cleaning and one post-cleaning recording are excluded from preprocessing, feature ranking, and model fitting in each split. Secondary purged blocked validation assessed temporal segments within recordings already represented during model development. In the primary record-level evaluation, the centred random forest correctly classified five of six consensus recording scores and achieved a record-level area under the receiver operating characteristic (ROC) curve (AUC) of 0.889. All eight classifiers achieved the same record-level accuracy in this dataset. In the secondary window-level analysis, random forest achieved 99.48% mean balanced accuracy, 99.45% pooled accuracy and macro-averaged F1 score (macro-F1), and an AUC of 1.000. However, the two acquisition groups are recorded on different dates, and their speed ranges did not overlap. A mean-speed-only control perfectly separated the window groups, showing that the residual-feature results cannot be interpreted independently of operating regime. The findings therefore demonstrate within-event acquisition-group discrimination, not operational cleaning-state diagnosis. Future validation requires matched operating conditions, repeated cleaning events, overlapping load and speed ranges, additional thermodynamic variables, and independent vessel-level testing.