Abstract
The purpose of this research is to apply machine learning methods to predict field reliability of household electromechanical appliances given properties of their prototypes. The scope of household electro-mechanical appliances is narrowed down to electric steam Irons. Research data is supplied by an appliances distributing organization referred to as Organization X for confidentiality reasons. The research approach involves data collection, data exploration, selection of a machine learning technique, model training, model performance evaluation, and performance improvement. Based on the nature of data available, resource economy, convenience, and accuracy requirements, a naïve Bayes algorithm is adopted for building the prediction model. The MATLAB fitcnb function is employed to develop the naïve Bayes prediction model. Probability distributions for predictor variables data are estimated using the MATLAB distribution fitter application. The identified distributions are used for further developing the naïve Bayes model. The developed naïve Bayes model is cross validated using the leave one out method and a prediction accuracy of 71% is obtained. After feature selection, the highest prediction accuracy of 78% is achieved. To evaluate the discrimination ability of the prediction model, Receiver Operating Characteristic (ROC) analysis is performed and an average Area Under Curve (AUC) of 0.86 is achieved. The research illustrates that machine learning is applicable for reliability prediction despite having a relatively small data. Furthermore, the work shows that there is a significant link between product design and reliability. The proposed method allows industry practitioners to evaluate reliability of new models of electro-mechanical appliances using a relatively small amount of data, in a timeous manner, and in a cost-effective way. The method presented solely utilizes the design and performance features of an appliance, for field reliability prediction.