Signed Wasserstein distance for granular process monitoring in Additive Friction Stir Deposition
Date Issued
May 1, 2026
Author(s)
Abir, Abrian Islam
The University of Tennessee
Advisor(s)
Brett G. Compton
Additional Advisor(s)
Bradley H. Jared
Sudarsanam S. Babu
Stian K. Romberg
Abstract
Additive Manufacturing (AM) processes, particularly solid-state systems like Additive Friction Stir Deposition (AFSD), display complex thermo-mechanical behavior that challenges conventional process monitoring strategies. Unlike fusion-based AM, AFSD relies on frictional heating and large plastic deformations, producing dynamically evolving microstructures without melting. The coupled thermal-mechanical physics, controller feedback loops, and sensitivity to environmental changes introduce process variations difficult to detect using traditional time-series metrics. This thesis introduces the one-dimensional Wasserstein distance (W_1) as a statistical dissimilarity metric for granular AFSD process monitoring. By comparing empirical distributions rather than time-aligned signals, the metric eliminates phase shift penalties common in Euclidean based approaches while retaining computational feasibility. Building on prior completed validation of unsigned Wasserstein metrics, this research extends the framework through a signed Wasserstein decomposition that separates dissimilarity into directional shift and dispersion components. This enables interpretation of whether deviations arise from mean operating point shifts (e.g., material hardening effects) or from increased variability (e.g., thermal instability, machine compliance). Experimental validation was conducted on a MELD L3 AFSD system using Aluminum 7075 feedstock. The unsigned metric successfully identified deviations exceeding a 0.1 threshold during thermal insulation and manufacturing pause studies. The signed extension displayed physically meaningful directionality: post-pause layers exhibited torque shifts consistent with age-hardening combined with startup effects, while insulation produced persistent mechanical loading redistribution without altering controller setpoint temperature. To evaluate cross-process generalizability, the framework was applied to Wire Arc Additive Manufacturing (WAAM) and Direct Ink Writing (DIW). Computational benchmarking confirms O(𝑛 log 𝑛) scaling for the Wasserstein distance calculations, supporting real-time deployment. This work transforms the Wasserstein distance from a scalar similarity measure into a diagnostic tool capable of linking statistical deviations to underlying physical mechanisms in additive manufacturing.
Degree
Master of Science
Major
Mechanical Engineering
File(s)![Thumbnail Image]()
Name
ABIR-PRIMARY-2026.pdf
Size
5.12 MB
Format
Adobe PDF
Checksum (MD5)
704c6bd0ee71ac37128011e9a8e50e69