Improving predictive maintenance classifiers of industrial sensors' data using entropy. A case study

The increase in the availability of sensors' data in manufacturing (Industrial Internet of Things, IIOT) poses the challenge on how best to use this information. One of the emerging applications of data analysis in this field is predictive maintenance: being able to identify when and why a certain component breaks down and empower early intervention to prevent breakdowns. Imbalanced datasets literature shows that tree models perform better with entropy splits than Gini index splits. Entropy measures applied in previous studies in the domain of industrial sensors' data include not only Shannon's but also Renyi and Tsallis. This paper looks at the performance of classifocation trees using different entropies applied to the Scania trucks dataset. In this case, the best performing tree is a C5:0 model but we confirm that Renyi and Tsallis entropy trees can improve classification of the minority class in the data without excessively penalising the classification of the majority class. These models can therefore help to prevent companies costs by improving the identification of possible failures and avoiding unnecessary interventions on well-working equipment