Session: 01-05-01: Applications of Artificial Intelligence/Machine Learning for Aerospace Structures 1

Paper Number: 152304

152304 - Whose Plane to Take? A Novel Approach to Identify the Optimal Model for Defect
Detection in Composite Aerostructure Material 

Carbon-fiber reinforced polymers are a common material used to manufacture aerospace structures due to their good tensile strength and resistance to compression. During the manufacturing process, these composites are thoroughly inspected for flaws and defects to ensure structural integrity during commercial use. Due to the strict safety standards in aerospace manufacturing, the inspection process is done manually and can significantly slow-down throughput. Recently, researchers developed DOC-DICAM, an AI-based assistance tool to identify defects in composite material. They showed their method can quickly and accurately identify defects in the fuselage section of the aircraft. Due to the repeated production of identical fuselages, the researchers were able to train and validate their models on archived fuselages.

In this work, we extend their method by identifying the optimal model that was trained with the archived fuselage closest to the new fuselage. We assume a collection of models exist with the same architecture each trained on different fuselages over time where any of these models can be used for the new fuselage. A naïve approach is to use the model trained on the most recent archived fuselage minimizing concept drift. However, we show that this may not always give the best results suggesting the need for a recommendation step in the workflow. While methods to measure the similarity between two probability distributions exists, the high volume and velocity of fuselage data means that archived data is routinely moved from disk to tape storage making the comparison between new and old data impossible.

Therefore, we propose a method to identify the optimal model using only the collection of models and the new data. We do this by incorporating an autoencoder into the training workflow that learns to reconstruct the original input. To recommend a model, we reconstruct the new data using the collection of archived models. The model that produces the lowest error is then used for the new fuselage. Since the DOC-DICAM workflow was developed as a multi-task learning framework, we only need to add a new task that reconstructs the input. Thus, a single model is capable of both reconstructing the input and identifying defects making the increase in complexity minimal. We validate our approach by showing the rankings produced by our method are consistent to those produced using conventional methods when archived data is available. Finally, we show that our method identifies the optimal model for identifying defects such as delmanination, porosity, and FOD while producing minimal false positives.

Presenting Author: Austin Yunker Argonne National Laboratory

Presenting Author Biography: Austin Yunker received a master's degree in statistics from Western Michigan University focusing on the statistical foundations of machine learning and artificial intelligence. He is continuing this work as a predoctoral appointee at Argonne National Laboratory in their Data Science and Learning Division. His work has focused on the development of AI/ML systems for science problems at scale using HPC resources including image restoration and reconstruction. Currently, he is working on designing an artificial intelligence system for improving non-destructive inspection methods used in evaluating state-of-the-art aircraft.