An integrated computer vision and machine learning system for emulsion processing /

Abstract

The existing quality evaluation of emulsions is typically based on subjective examination of samples under the microscope by trained analysts. The major drawbacks of such manual assessment include inter-observer variability, intra-observer variability, lack of speed, poor accuracy and it is also prone to over-processing. Other conventional droplet analysis techniques such as laser diffraction and spectroscopy, which require timeconsuming sample preparation, have been verified as unreliable and introduce an additional complexity to industrial processes. In order to overcome these challenges, a novel automated approach based on image segmentation and machine learning is investigated in this research for the quality evaluation and optimisation of industrial emulsion processing. Bright field micrographs were obtained during an industrial emulsification process. Two image segmentation techniques, Edge & Symmetry (EST) and Histogram-Based (HBT), were applied to detect the oil droplets from the micrographs. These techniques were also used to extract various morphological characteristics of the droplets. The most significant predictors were selected from these droplet characteristics for developing machine learning models. The most efficient image segmentation technique was also identified. The micrographs were grouped into four quality-based categories identified as TAMU (Target, Acceptable, Marginal and Unacceptable). Supervised machine learning and deep learning models were developed for the TAMU classification of unknown emulsion micrographs. A comparative study was performed between manual and machine learning classification using Attribute Agreement Analysis. Regression models were developed to predict the RPT (Remaining Processing Time) required, at all stages of emulsification, to achieve the target characteristics. These prediction models were intended to avoid over-processing in emulsion manufacturing. HBT exhibited excellent potential in droplet detection and characterisation compared to the EST approach. HBT was successful in detecting droplets with diameter as low as ca. 1 µm from emulsion samples having dispersed phase fraction ≈ 50%. The machine learning classification models presented high accuracies ranging from 92% to 100%. The deep learning models demonstrated lower accuracies from 44% to 89%. The results of the comparative analysis showed that the machine learning classification is superior to manual classification with respect to speed (180 times faster), greater accuracy (10% to 40%) and repeatability. The prediction models presented an adjusted R2 ≈ 92%. The entire automated approach based on image segmentation and machine learning was implemented as a soft sensor. The soft sensor supports the real-time deployment of the technique into an industrial environment. The proposed approach has the potential to predict instantaneous product quality as well as the process time required to achieve the desirable droplet characteristics. This will avoid over-processing and wastage of resources leading to more efficient and sustainable emulsion manufacturing.