Machine Learning for Combined Fault Diagnosis in Rotating Machines From Training With Simple Faults

This paper presents a machine learning framework for diagnosing simple and combined mechanical faults in rotating machines from vibration data. The main goal is to investigate whether a classifier trained mostly on simple faults can generalize to previously unseen fault combinations. The study uses the COMFAULDA dataset, whose instances are represented by 217 features extracted from the time and frequency domains. Four classifiers are evaluated under a binary multi-target formulation: multilayer perceptron, Random Forest, XGBoost, and support vector machine. Five experimental scenarios are defined, from zero exposure to combined faults during training to progressive partial exposure. Results show that pure compositional generalization is difficult, but controlled inclusion of a small fraction of combined-fault samples significantly improves performance. The best result was obtained by XGBoost with Bayesian hyperparameter optimization, achieving a combined \(F_1\) score of 0.4489 and an overall accuracy of 0.8796 in the most favorable scenario. The results indicate that the proposed framework is promising for predictive maintenance applications in which combined faults are rare or only partially labeled.