DEPOSITION QUALITY OPTIMIZATION OF ADDITIVE FRICTION STIR DEPOSITED ALUMINIUM ALLOY USING UNSUPERVISED MACHINE LEARNING

Abstract

Additive friction stir deposition (AFSD) is a promising solid-state additive manufacturing technology, but achieving continuous high deposition quality remains challenging due to complex process-structure connections. This study investigates unsupervised machine learning algorithms for mapping process parameters to deposition outcomes without requiring extensive labelled data. On experimental deposition data, algorithms including hierarchical clustering, k-means, spectral clustering, Gaussian mixtures, autoencoders, and self-organizing maps are used. The algorithms find intrinsic patterns and groupings in the multi- factor process data in an unbiased manner. With a silhouette score of 0.7618, k-means clustering performed the best, showing cohesive data clustering. Visualizations like dendrograms and trained maps shed light on the links between process parameters and deposition quality. The cluster analysis identifies process conditions that result in poor deposition quality. This highlights the ability of unsupervised approaches to capture deposition process patterns based solely on data with no prior system knowledge. This data-driven strategy has the potential to significantly improve AFSD process optimization and control, with implications for boosting industrial adoption. The unsupervised learning framework lays the groundwork for using process data to improve the quality and productivity of advanced manufacturing procedures.

Keywords

• additive friction stir deposition
• unsupervised machine learning
• deposition quality
• silhouette score