Download PDF: A Combination of Techniques Leads to Improved Friction Stir Welding The NESC developed several innovative tools and techniques during an assessment to find the root cause of poor tensile strength and low topography anomalies (LTA) in welds formed using a solid-state welding process called self-reacting friction stir welding (SRFSW).
Using a combination of machine learning, statistical modeling, and physics-based simulations, the assessment team helped improve the weld process and solve both issues, lifting constraints that had been placed on flight hardware.
Developing Techniques for LTA Detection
Determining the root cause of poor tensile strength welds and LTA observed on the weld fracture surfaces involved several techniques:
Deep Learning for LTA Detection: The NESC team developed a machine-learning model to detect and segment LTA in weld images. The model was trained on images annotated by metallurgy experts, with a majority-vote consensus to resolve disagreements. The team then developed an accompanying standard operating procedure for image capture to improve robustness and reduce bias. This model was built on previous NASA work to develop specialty microscopy analysis foundation models by pretraining on 100,000+ microscopy images. This step was crucial to linking process parameters with LTA occurrence in an objective, nonbiased way.
Integrated Data-Ingestion Framework: SRFSW is a complex process with many interacting variables. The weld process produces a large amount of data with diverse data types that include dozens of tabular process parameters, dozens of sequential data streams from the production tool, fracture and weld cross-section images, and mechani …