Predicting bilgewater emulsion stability by oil separation using image processing and machine learning. (1st September 2022)

• Record Type: Journal Article
• Authors: Lee, Woo Hyoung; Park, Cheol Young; Diaz, Daniela; Rodriguez, Kelsey L.; Chung, Jongik; Church, Jared; Willner, Marjorie R.; Lundin, Jeffrey G.; Paynter, Danielle M.
• Is Part Of: Water research. Volume 223(2022)
• Publication Date: 2022-09-01
• Subjects: Bilgewater -- Coalescence -- Emulsion stability -- Image processing -- Machine learning -- Oil-in-water emulsions; Water -- Pollution -- Research -- Periodicals; 363.7394
• DOI: 10.1016/j.watres.2022.118977
• Languages: English
• ISSNs: 0043-1354

Abstract

Highlights: A novel image processing was conducted to determine oil separation levels. Predictive models for bilgewater emulsion stability were first developed. Coalescence and kinetic emulsion stabilities were separately determined. Abstract: Bilgewater is a shipboard multi-component oily wastewater, combining numerous wastewater sources. A better understanding of bilgewater emulsions is required for proper wastewater management to meet discharge regulations. In this study, we developed 360 emulsion samples based on commonly used Navy cleaner data and previous bilgewater composition studies. Oil value (OV) was obtained from image analysis of oil/creaming layer and validated by oil separation (OS) which was experimentally determined using a gravimetric method. OV (%) showed good agreement with OS (%), indicating that a simple image-based parameter can be used for emulsion stability prediction model development. An ANOVA analysis was conducted of the five variables ( Cleaner, Salinity, Suspended Solids [ SS ], pH, and Temperature ) that significantly impacted estimates of OV, finding that the Cleaner, Salinity, and SS variables were statistically significant ( p < 0.05), while pH and Temperature were not. In general, most cleaners showed improved oil separation with salt additions. Novel machine learning (ML)-based predictive models of both classification and regression for bilgewater emulsion stability were then developed using OV. For classification, the random forest (RF) classifiers achieved the most accurate prediction with F1-score of 0.8224, while in regression-based models the decision tree (DT) regressor showed the highest prediction of emulsion stability with the average mean absolute error (MAE) of 0.1611. Turbidity also showed a good emulsion prediction with RF regressor (MAE of 0.0559) and RF classifier (F1-score of 0.9338). One predictor variable removal test showed that Salinity, SS, and Temperature are the most impactful variables in the developed models. This is the first study to use image processing and machine learning for the prediction of oil separation for the application of bilgewater assessment within the marine sector. Graphical abstract: Image, graphical abstract