Comparison of Artificial Intelligence Methods for Prediction of Mechanical Properties

Abstract

This paper compares artificial intelligence (AI) methods to predict mechanical properties of sheet metal in stamping processes. The deviation of the mechanical properties of each blank leads to unpredicted failures in stamping processes, such as fracture and spring back. The research team of this paper has been building a real time control system for stamping process in a smart factory. In order to facilitate that, it is necessary to predict the mechanical properties of each blank with non-destructive testing. The regression models based on the linear algebraic scheme have traditionally brought reliable results in terms of matching the measured non-destructive testing values to the mechanical properties. With a parallel to algebraic regression models, in recent studies on various domains, AI models have been adopted to improve the accuracy of the end-results and effectiveness of the models. This paper discusses the applicability of AI models for predicting the mechanical properties based on the eddy-current non-destructive testing method. For the study, 6 input features are collected through the eddy-current non-destructive testing to map eddy-current input data to mechanical properties of the blank. Yield stress and uniform elongation were predicted by using five AI methods, i.e., regularized linear regression, support vector regularized linear regression, support vector regression, multi-layer neural network, random forest regression, and gradient boosting regression were compared. The model performance, validated with 20% of test data that are intact during the training phase, is the main discussion point of this paper. Future works to improve the predictive accuracy of AI models is also discussed.

Monitoring of fluctuating material properties for optimizing sheet-metal forming processes: a systematic literature review

“…Heingärtner used multidimensional regression based on eddy current data to calculate the material parameters tensile strength, yield strength, uniform strain, and elongation at break in inline production [7]. Moreover, Lee et al tested different machine learning algorithms to predict the yield stress and the uniform elongation based on 6 features of eddy current signals and achieved a high accuracy [16]. Zoesch et al were able to detect cracks and thinning of material during deep drawing with two sensors (eddy current, 3MA) [17].…”

“…Sheet-metal forming in general [1,3,4,20] RQ1 Magnetic barkhausen noise [9][10][11][12] Eddy current [6,7,16,17] 3MA [13][14][15] IMPOC [18,19]…”

A Study on Prediction Performance Measurement of Automated Machine Learning

“…While the gradient boosting model (Friedman, 2002) can reduce the vias of the machine learning model, the problem of overfitting that can occur in the learning structure occurs. To solve this problem, XGB tried to overcome the problems of GBM by applying a penalty to the weights (Lee et al, 2020) by reflecting regularization and subsampling.…”

“…Random Forest (Breiman, 2019) is a machine learning method in which an individual classification and regression tree (CART) combines several decision trees (Lee et al, 2020). It is one of the representative techniques of bagging (Bootstrap Aggregation), which performs aggregation by resampling multiple data sets through boosting.…”

The influence of aging on buckle strength loss in AA5182-H48 for beverage can ends

“…Artificial neural network-based prediction models for aluminum alloys have gained significant traction and widespread industrial application in recent years. Numerous studies have focused on developing prediction models for mechanical properties and manufacturing processes (Alibeiki et al, 2012;Lee et al, 2020;Wang et al, 2015;Zhang et al, 2019). For example, a paper by Merayo et al (2020) introduced a computer-aided tool that predicts the plastic behavior, yield strength, and ultimate tensile strength of metallic materials.…”