Determination of Forming Limits in Sheet Metal Forming Using Deep Learning

Area of Science:

• Materials Science
• Mechanical Engineering
• Computational Mechanics

Background:

• The forming limit curve (FLC) is crucial for predicting sheet metal instability in forming processes, often determined via optical strain analysis during Nakajima tests.
• Existing methods like DIN EN ISO 12004-2 and heuristic time-dependent approaches have limitations, particularly with brittle materials lacking distinct necking phases.
• Recent pattern recognition methods, while advanced, still rely on prior knowledge, time, and localization data.

Purpose of the Study:

• To develop a novel, data-driven method for determining the forming limit curve (FLC) that overcomes the limitations of current techniques.
• To enable accurate instability prediction in sheet metal forming, even for challenging materials and incomplete test data.
• To establish a location and time-independent approach for FLC determination using advanced machine learning.

Main Methods:

• A Siamese convolutional neural network (CNN) was employed as a feature extractor to automatically learn relevant features from forming data.
• A supervised setup was used to train the CNN on the distinct homogeneous and inhomogeneous forming phases.
• Unsupervised clustering using Student's t mixture models was applied to the learned features, categorizing the forming process into three distributions.

Main Results:

• The developed method automatically learns features and clusters the complete forming process into three distributions, independent of location and time.
• Knowledge gained from fully formed specimens can be transferred to prematurely stopped forming processes, enabling probabilistic assessments.
• The method demonstrated generalization capabilities across different materials, including an aluminum alloy exhibiting Portevin-LE Chatelier effects.

Conclusions:

• The Siamese CNN combined with Student's t mixture models provides a robust and versatile approach for FLC determination, overcoming limitations of traditional and recent methods.
• This technique enhances the analysis of sheet metal instability, particularly for brittle materials and scenarios with incomplete strain data.
• The location and time-independent nature allows for broader application and knowledge transfer across various forming processes and materials.