Estimation of Material Properties Using Machine Learning

Abstract

The aim of this thesis was to study and compare different machine learning regression methodologies to predict the material properties using Barkhausen noise measurements. Some of the features of Barkhausen noise signal for training the model included, for example, root mean square (RMS) value, peak height, peak position, peak width, coercivity. The measurements were taken at the laboratory of Material Science and Environmental Engineering at Tampere University and the dataset consisted of small, middle and large sized ground components. This dataset was further divided into subsets based on the sizes of the components and removing the measurements of hard turned spots. The whole process is divided into four steps, feature generation, feature selection, model training and model testing. Feature generation includes the measurements and pre-processing steps to create the datasets. The most significant features are selected using recursive feature elimination (RFE) technique based on the correlation between features and target variables. In the model training and testing step, different machine learning regression algorithms were fitted and tested on the datasets. The tested methods included linear regression, ridge regression, LASSO regression, elastic net regression, random forest regression and gradient boosting regression. Random forest regression had the best performance on average among the other methods on all the five sets of data.

The overall performances of models improved with the recursive feature elimination (RFE) method. But these changes were very minor for the linear models and quite impressive for the decision tree models. The overall test results were very good with an error of less than 25% in worst cases and less than 2% in best cases. These results were further improved when the models were trained and tested with the measurements from separate sized specimen, divided based on the diameter. The error in this case was less than 20% for worst cases and less than 1% for best cases.