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 the residual strength of CFRP laminates based on machine learning. A diverse dataset is acquired 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, preprocessed 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 laminates. The effectiveness of the current method is confirmed, with the mean Intersection over Union (mIoU) achieving over 93% for damage identification, the Mean Absolute Error (MAE) reducing to 5.4% for damage depth prediction, and the Mean Relative Error (MRE) reducing to 7.6% for residual strength prediction, respectively.
Keywords: carbon fiber-reinforced polymer; composites; identification; lightning strike damage; machine learning; prediction; residual strength.