engleski

Application of feature selection techniques in assessing variables relevant for estimation of materials parameters and behavior

To adequately design a structure or a part, the behavior of a material, more precisely, relation between its stresses and strains, must be known. Considering that the experimental determination of fatigue and cyclic material parameters is costly and long-lasting, as opposed to monotonic tensile tests, it is of great interest to use more easily obtainable monotonic properties to estimate cyclic and fatigue material behavior [1]. Building a predictive model from acquired data can be done using classical approaches, such as regression, or more recently, various available machine learning methods. The dataset which is used as an input for such models needs to be of appropriate size and have an adequate number of input variables, also called predictors, to avoid underfitting or overfitting a model. Higher ratio of number of samples to number of predictors makes the model less likely to be affected by possible errors in data and to generalize new cases well [2]. To increase the data volume, additional datasets can be acquired by performing experiments which take a great amount of time. Therefore, it is more economical to implement a feature selection (i.e. feature engineering) techniques instead, which enable the detection of redundant input variables followed by their removal which then consequently reduces model complexity [2]. In this study, using several chosen feature selection techniques, importance of each predictor in relation with the response is determined and a subset of the most relevant variables for predicting the cyclic and fatigue parameters is selected. Building models with newly acquired subset should reduce overfitting, improve interpretability and decrease the complexity of the model.

Material behavior ; Cyclic/fatigue parameters ; Feature engineering/selection

nije evidentirano