Topological Data Analysis Based Feature Selection for Predicting Fatigue Strength of Steel Using Machine Learning

Abstract

Routine Machine learning methodologies are based on Statistics and traditional data analysis methods, which may account for loss of some important information from the data during analytical process. So we need a new set of tools and algorithms that can effectively make most of such information. This paper investigates the implementation of new and growing data analytics technology known as topological data analysis (TDA). We use Mapper algorithm from TDA on data taken from public domain database of National Institute for Material Science (NIMS), for selecting key process parameters or features that impacts fatigue strength of steel. Mapper algorithm outputs a network graph called topological network that captures the important connectivity information and complex interactions between the clusters present in the data where traditional methods fail to capture such interactions. We analyse the shape of topological network obtained from mapper for selecting important features that impact the fatigue strength of steel. We then use XG-Boost prediction model with tuned hyper parameters using various tuning algorithms, to evaluate the impact of the selected features. This approach of data analysis and modelling helps to monitor and control the process in a more cost-effective manner. The experimental results have successfully proved the efficacy of TDA and machine learning (ML) and hence developed the predictive model for the same.