Multi-task learning application for predicting impact damage-related information using surface profiles of CFRP laminates

Impact damage prediction has been considered a critical issue for several years, especially in manufacturing or maintenance. Several researchers have been studying on impact detection or damage prediction on composite materials applying machine learning, a data driven analysis methodology. This study develops the decision tree based multi-task learning scheme for the prediction of impact damage information solely from an external surface profile. Multi-task learning enables effective learning; in other words, it can integrate the relationships among objective variables. Low-velocity impact tests and damage measurement were conducted to create the dataset and investigate the correlations between the impact damage and impact conditions. Using the features designed from the surface profile data, multi-task learning was applied to predict the impactor shape and delamination extent. By comparing the effectiveness of the proposed method and that of the original single -task learning method, it was inferred that the multi-task learning has advantages in the prediction accuracy and model plausibility, considering the impact phenomenon.

Automated summary

This study develops a decision tree based multi-task learning scheme to predict impact damage information solely from an external surface profile. Through low-velocity impact tests and damage measurement, a dataset is created to investigate the correlations between impact damage and impact conditions. The results infer that multi-task learning has advantages in prediction accuracy and model plausibility.