Fast evaluation of crack growth path using time series forecasting. (September 2019)

• Record Type: Journal Article
• Authors: Do, Dieu T.T.; Lee, Jaehong; Nguyen-Xuan, H.
• Is Part Of: Engineering fracture mechanics. Volume 218(2019)
• Publication Date: 2019-09
• Subjects: Time series forecasting -- Crack propagation -- Damage mechanics -- Machine learning -- Long short-term memory -- Multi-layer neural network; Fracture mechanics -- Periodicals; Rupture, Mécanique de la -- Périodiques; Fracture mechanics; Periodicals; 620.112605
• DOI: 10.1016/j.engfracmech.2019.106567
• Languages: English
• ISSNs: 0013-7944

Abstract

Highlights: Forecasting the crack propagation in risk assessment of engineering structures is based on long short-term memory and multi-layer neural network algorithms. The present approach helps reducing significantly computational cost without any analysis tools. Only a small amount of data from numerical analysis or experiment processes is required. The state-of-the-art techniques such as dropout and mini-batch are utilized in the proposed method to enhance computational performance. The most appropriate architecture is proposed to ensure accuracy, stability and low computational cost. Abstract: This paper aims at forecasting the crack propagation in risk assessment of engineering structures based on time series algorithms named "long short-term memory" and "multi-layer neural network". The core idea is how to predict precisely the accuracy of solution of crack growth in engineering fracture structures without requiring re-modeling and re-computational attempts. The underlying method only requires a small amount of data from numerical analysis or experimental processes. Based on optimal parameters learned from information of a given dataset, the machine learning methods are used to quickly forecast the crack growth without any analysis tools. In addition, the advanced techniques such as dropout and mini-batch are also utilized to enhance computational performance. The effectiveness and accuracy of the current method are verified by comparing gained results with those from numerical approaches or experimental data.