Dynamic prediction of aluminum alloy fatigue crack growth rate based on class incremental learning and multi-dimensional variational autoencoder

Abstract

Aluminum alloys, valued for their low density and high strength-to-weight ratio, are crucial in the aerospace industry. To enhance their security and maintenance economy in service, this work employs a class incremental learning method (CIL) to predict the fatigue crack growth rate of 2xxx and 7xxx series aluminum alloys. The developed multi-dimensional autoencoder class incremental learning and data update monitoring feature enhanced prediction model (MDACIL-RTM-FIEP) integrates mechanical, environmental, and material features using variational autoencoders (VAE) for deep feature learning. Results show that the model, through the data update monitoring and incremental triggering mechanism (DUM-IT), adapts effectively to data changes, significantly enhancing prediction accuracy. The model’s fatigue crack growth rate (FCGR) incremental learning accuracy (IAC) improved to 0.9644 from 0.8715, and its backward transfer (BT) value decreased to −0.0182 from −0.6325, indicating excellent adaptability of new knowledge and retention of old knowledge. It also addresses the “catastrophic forgetting” issue, underscoring the effectiveness of CIL strategies in dynamic data environments.