Failures Prediction in Non-Maintainable Components for Industrial Maintenance Management Using Machine Learning Techniques

Predictive maintenance of non-maintainable industrial components poses a significant challenge in maintenance management due to the complexity of predicting imminent failures at early stages, influenced by various factors affecting the operation of industrial assets. This paper proposes a new methodology that develops supervised classification Machine Learning (ML) models to predict failures in non-maintainable components by utilizing Time in Service (TIS) as a key feature in the context of a centrifugal pump, a critical asset in industrial processes. Two algorithms are evaluated: Random Forests (RF) and K-Nearest Neighbors (K-NN), using a dataset of 2,000 historical maintenance records. Data partitioning is done through a combination of k-Fold Cross-Validation and Grid Search, and hyperparameters are optimized by minimizing the loss function. The evaluation metrics (precision, recall, accuracy, and F1-Score) are derived from confusion matrices. Results indicated that while both models exhibited competitive performance, RF models achieved a higher average recall (97.11%) than K-NN models (95.2%), which is critical for effective failure detection. This study confirms the feasibility of using supervised ML models to predict failures in non-maintainable industrial assets, with TIS as a foundational feature. The developed models provide a valuable tool for enhancing operational reliability, optimizing maintenance management, and facilitating data-driven decision-making, all in compliance with ISO 14224:2016 standards.