Version 1
: Received: 1 June 2022 / Approved: 2 June 2022 / Online: 2 June 2022 (11:08:35 CEST)
Version 2
: Received: 18 August 2022 / Approved: 19 August 2022 / Online: 19 August 2022 (06:03:08 CEST)
How to cite:
Matlala, K.; Mishra, A. K.; Puthal, D. Detection and Analysis of Molten Aluminium Cleanliness Using a Pulsed Ultrasound System. Preprints2022, 2022060033. https://doi.org/10.20944/preprints202206.0033.v1
Matlala, K.; Mishra, A. K.; Puthal, D. Detection and Analysis of Molten Aluminium Cleanliness Using a Pulsed Ultrasound System. Preprints 2022, 2022060033. https://doi.org/10.20944/preprints202206.0033.v1
Matlala, K.; Mishra, A. K.; Puthal, D. Detection and Analysis of Molten Aluminium Cleanliness Using a Pulsed Ultrasound System. Preprints2022, 2022060033. https://doi.org/10.20944/preprints202206.0033.v1
APA Style
Matlala, K., Mishra, A. K., & Puthal, D. (2022). Detection and Analysis of Molten Aluminium Cleanliness Using a Pulsed Ultrasound System. Preprints. https://doi.org/10.20944/preprints202206.0033.v1
Chicago/Turabian Style
Matlala, K., Amit Kumar Mishra and Deepak Puthal. 2022 "Detection and Analysis of Molten Aluminium Cleanliness Using a Pulsed Ultrasound System" Preprints. https://doi.org/10.20944/preprints202206.0033.v1
Abstract
The analysis of data produced by the MV20/20 sensor, tagged with quality outcomes, is presented with the aim of developing a predictor model for real-time anomaly detection and classification. Three types of inclusions, undesired particles that deteriorate the quality of production, are used to tag the quality data using results from the lab. We explore both unsupervised and supervised learning, which both offer advantages in monitoring and controlling the quality of production. It is discovered that the dataset can be clustered using techniques like K-Means and DBSCAN. Bounding the data within a 95% confidence interval ellipse ensures we can detect anomalous events in real time. For supervised learning, a two-stage classifier is explored, which classifies the outcome of a cast and secondly the inclusion responsible for the negative outcome. We explore models from logistic regression and support vector machines, to two neural networks, namely the multi-layer perceptron and the radial basis function network. While the cast outcome is adequately predicted by all the models, the multi-layer perceptron provides a boundary performance for the inclusion type. A more advanced technique for model optimisation, namely grid search, is applied in order to improve on the results. The outcome for the grid search is not much better, which indicates a global maximum in the learning capacity of the model. Recommendations include the addition of sensor systems and an audit of data collection variation.
Keywords
MV20/20; PoDFA; anomaly detection; statistical process control; principal components analysis; K-Means; DBSCAN; multi-layer perceptron; inclusion; receiver operating characteristic; confusion matrix
Subject
Engineering, Industrial and Manufacturing Engineering
Copyright:
This is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
