Condition monitoring of fans under variable-speed operation is challenging because changes in vibration characteristics can reduce the generalizability of machine-learning models. This study presents a machine-learning framework for identifying three fan imbalance configurations under previously unseen operating-speed conditions. A cross-speed evaluation strategy based on leave-one-speed-out validation was employed to provide a more realistic assessment of classifier robustness compared with conventional random data splitting. Four machine-learning models, including Logistic Regression, RBF-SVM, Random Forest, and Extra Trees, were evaluated using engineered features extracted from triaxial vibration measurements. In addition, predictive entropy and conformal prediction were incorporated to assess the reliability and uncertainty of model decisions. Extra Trees achieved the highest balanced accuracy under cross-speed validation, while uncertainty analysis demonstrated the ability of reliability metrics to identify ambiguous predictions and support selective decision-making. The results show that validation strategy strongly influences the estimated performance of vibration-based classifiers and that reliability-aware evaluation provides additional information beyond accuracy alone. The proposed framework provides a realistic assessment of cross-speed generalization for the investigated experimental benchmark; however, further validation using independent equipment, sensing configurations, and operating conditions is required before broader industrial conclusions can be drawn.