Ultrasonic measurements and machine learning for monitoring the removal of surface fouling during clean-in-place processes

• Published in: Food and bioproducts processing 2020-09, Vol.123, p.1-13
• Main Authors: Escrig, J. Escrig, Simeone, A., Woolley, E., Rangappa, S., Rady, A., Watson, N.J.
• Format: Article
• Language: English
• Subjects: Algorithms; Clean-in-Place; Cleaning; Data; Data points; Datasets; Economics; Food; Food processing industry; Foods; Fouling; Industrial digital technologies; Learning algorithms; Machine learning; Manufacturing industry; Methods; Pipework; Training; Ultrasonic measurements; Ultrasonic technology
• ISSN: 0960-3085; 1744-3571
• DOI: 10.1016/j.fbp.2020.05.003

•Ultrasonic measurements used to monitor surface cleaning of food fouling materials.•Machine learning classification models developed to analyse ultrasonic signals.•Models performance best when input data extracted directly from ultrasonic waves.•Model performance depends on the fouling material been monitored.•Data combined from all fouling materials creates a better training data set. Cleaning is an essential operation in the food and drink manufacturing sector, although it comes with significant economic and environmental costs. Cleaning is generally performed using autonomous Clean-in-Place (CIP) processes, which often over-clean, as suitable technologies do not exist to determine when fouling has been removed from the internal surfaces of processing equipment. This research combines ultrasonic measurements and machine learning methods to determine when fouling has been removed from a test section of pipework for a range of different food materials. The results show that the proposed methodology is successful in predicting when fouling is present on the test section with accuracies up to 99% for the range of different machine learning algorithms studied. Various aspects relating to the training data set and input data selection were studied to determine their effect on the performance of the different machine learning methods studied. It was found that the classification models performed better when data points were extracted directly from the ultrasonic waves and when data sets were combined for different fouling materials.