Code for predicting the residual strength of CFRP laminates subjected to lightning strikes using a machine learning approach.
Description
Due to the complex and uncertain physics of lightning strike on carbon fiber-reinforced polymer (CFRP) laminates, conventional numerical simulation methods for assessing the residual strength of lightning-damaged CFRP laminates are highly time-consuming and far from pretty. To overcome these challenges, this study proposes a new prediction method for residual strength of CFRP laminates based on machine learning. A diverse dataset is gained and augmented from photographs of lightning strike damage areas, C-scan images, mechanical performance data, layup details, and lightning current parameters. Original lightning strike images, preprocesses with the Sobel operator for edge enhancement, are fed into a UNet neural network using four channels to detect damaged areas. These identified areas, along with lightning parameters and layup details, are inputs for a neural network predicting the damage depth in CFRP laminates. Due to its close relation to residual strength, damage depth is then used to estimate the residual strength of lightning damaged CFRP laminate. The effectiveness of current method is confirmed with the mean Intersection over Union (mIoU) achieving over 93% for damage identification, the Mean Absolute Error (MAE) and the Mean Relative Error (MRE) reducing to 5.4% for damage depth prediction and 7.6% for residual strength prediction, respectively.