Selecting the optimal supervised learning algorithm for fault diagnosis in predictive maintenance

Predictive Maintenance (PdM) plays a pivotal role in ensuring the longevity and reliability of machinery. As the landscape of PdM evolves with new data sets, changing scenarios, and unforeseen challenges, machine learning (ML) algorithms need to adapt accordingly. However, the performance of ML algorithms can fluctuate due to these dynamic factors, complicating the decision-making process for PdM applications. This study aims to address this challenge by conducting a comprehensive comparison of widely used ML algorithms for PdM, specifically focusing on identifying critical factors that determine the success of a model in real-world maintenance environments. The motivation behind this work stems from the need to propose a selection process that considers both the specific aspects of maintenance and the problem's specifications, particularly its objectives. Initially, we evaluated the algorithms based on the evaluation criteria extracted from the literature. However, we realized that certain criteria should carry more weight depending on the maintenance objectives. This is what we call the weighted voting method. As a result, we refined our approach by proposing a second round of algorithm selection, using the weighted voting method. To validate the effectiveness of this process, we applied our methodology to two distinct maintenance scenarios, demonstrating the enhanced efficiency and accuracy of the algorithm selection process.