Inferring Mechanical Properties of Wire Rods via Transfer Learning Using Pre-Trained Neural Networks

The primary objective of this study is to explore how machine learning techniques can be incorporated into the analysis of material deformation. Neural network algorithms are applied to the study of mechanical properties of wire rods subjected to cold plastic deformations. Specifically, this study explores how pre-trained neural networks with appropriate architecture can be exploited to predict apparently distinct but internally related features. Tentative predictions are made by observing only an insignificant cropped fraction of the material’s profile. The neural network models are trained and calibrated using 6400 image fractions with a resolution of 120×90 pixels. Different architectures are developed with a focus on two particular aspects. Firstly, different possible architectures are compared, particularly between multi-output and multi-label convolutional neural networks (CNNs). Moreover, a hybrid model is employed, essentially a conjunction of a CNN with a multi-layer perceptron (MLP). The neural network’s input constitutes combined numerical and visual data, and its architecture primarily consists of seven dense layers and eight convolutional layers. By proper calibration and fine-tuning, observed improvements over the standard CNN models are reflected by good training and test accuracies in order to predict the material’s mechanical properties, with efficiency demonstrated by the loss function’s rapid convergence. Secondly, the role of the pre-training process is investigated. The obtained CNN-MLP model can inherit the learning from a pre-trained multi-label CNN, initially developed for distinct features such as localization and number of passes. It is demonstrated that the pre-training effectively accelerates the learning process for the target feature. Therefore, it is concluded that appropriate architecture design and pre-training are essential for applying machine learning techniques to realistic problems.