MACHINE LEARNING PREDICTION MODEL FOR SMALL DATA SETS INSTEAD OF DESTRUCTIVE TESTS FOR A CASE OF RESISTANCE BRAZING PROCESS VERIFICATION

Abstract

This paper presents a case study of Machine Learning (ML) prediction model for small data sets instead of destructive testing of brazed contacts. The main problems noted in the study were data availability, data quality, an extremely low number of NOK destructive test results and overall small data set. Recent researches are not very often focused on small data set ML prediction models and even less often on its application in resistance brazing. This paper tends to bridge this gap. The case study methodology consists of data collection, data preparation, correlation analysis, feature selection, model training, hyperparameter optimization, and model evaluation. It is proven possible to train ML prediction model with small datasets to predict numerical test outcomes if dataset quality is adequate. The practical use of this approach is reflected in the reduction of test costs since destructive tests can be quite expensive, and ML prediction model is one time, relatively low investment.