AI/ML Applications to Identify Tank Cleaning Operations & Quantify Slop Discharge

Abstract

Chemical tankers discharge contaminated washwater from tank-cleaning operations under conditions regulated by MARPOL Annex II and the IBC Code, yet compliance cannot generally be verified externally at the time of discharge. Vessels are not required to report when, where or in what volume tank cleaning occurs and current monitoring relies on aerial surveillance that is constrained by weather, daylight and geographic coverage. This leaves a gap in the environmental oversight of chemical tanker traffic in sensitive regions such as the Baltic Sea. Crucially, no reliable labelled record of confirmed tank-cleaning events exists, which precludes supervised machine learning approaches. This thesis develops and evaluates a reproducible, label-free pipeline for identifying vessel trajectories whose navigational behaviour is potentially consistent with tank cleaning activity, using historical AIS data. The work is conducted in collaboration with Scanjet AB and is based on AIS records for the Baltic Sea from January to March 2021. Segmented vessel trajectories are enriched with engineered behavioural features; speed variation, turning-angle variation and drift, and compressed before being rendered as multi-channel image tensors. A self-supervised representation learning framework (BYOL) with a custom convolutional encoder learns compact behavioural embeddings from these tensors without labels, and HDBSCAN cluster ing is applied to isolate behaviourally uncommon trajectories for expert-informed interpretation. Two trajectory compression methods, DP and TDKC, are system atically compared and the contribution of the engineered behavioural features is assessed through ablation. Applied to 91,431 learned trajectory embeddings, the framework organised vessel movement into a small number of coherent behavioural clusters while isolating a mi nority of sparse, behaviourally distinct trajectories consistent with irregular manoeu vring, looping and offshore-deviation patterns. Behavioural feature engineering was found to be decisive: removing the drift channel fragmented the embedding space and increased the proportion of unassignable trajectories from 36% to 69%. The compression comparison showed that, in the configurations tested, DP supported more coherent representation learning than TDKC; this difference is attributable primarily to the amount of trajectory structure retained during compression rather than to the compression criterion itself. Detected anomalies were combined with operational washing-system metadata from Scanjet AB to produce scenario-based estimates of potential washwater discharge volumes. The pipeline is presented as a detection-support tool rather than an autonomous classifier: its outputs are candidate trajectories requiring expert review, not con firmed evidence of discharge. The thesis demonstrates that self-supervised repre sentation learning provides a viable foundation for large-scale behavioural anomaly analysis in label-scarce maritime monitoring contexts and offers a transferable basis for future environmental-monitoring systems.