Toward Using Few-Shot Learning for Prediction of Complex In-Service Defects of Composite Products: A case Study

Numerical modelings for industrial trials, which are used to predict expensive defects, are unrealistic. As a result, since generating annotated data is very expensive in the industry, a model is needed that can be adapted to real-world parameters with a handful of real experiments. Thus, this paper investigates applying the model-agnostic meta-learning algorithm, which is one of the state-of-the-art models in few-shot learning, to create a machine learning model for predicting bunching defects in manufacturing industrial hoses. This model is expected to extract the knowledge from simulated data from the bunching defects and adapt this knowledge for the real world using a few examples collected from real experiments. While the accuracy of the proposed algorithm is less than 15 percent, this paper shows a promising result if better-simulated data can be used.