Drift-Aware Online Ensemble Learning for Imbalanced Industrial Failure Detection: A Comparative Study

Abstract: Predictive-maintenance models are usually reported on fixed, shuffled train-test splits. This hides two practical weaknesses: class imbalance, which makes accuracy misleading, and sensor drift, which static models never meet during evaluation. This paper presents a comparative study on the public AI4I 2020 dataset (10,000 records, 3.39% failures) that addresses both. First, we repeat the 80/20 evaluation protocol of a recent benchmark study. Gradient boosting with physics-informed features (machine power, temperature difference and wear-torque load) reaches a failure-class F1-score of 0.852 (precision 0.979, recall 0.754, ROC-AUC 0.979), against 0.823 published for the benchmark's tuned XGBoost model. The bootstrap 95% interval [0.774, 0.917] contains the published value, so the gain is indicative rather than conclusive. Second, we simulate sensor drift at five severities and compare static models with online learners that retrain on a sliding window, triggered by the Drift Detection Method (DDM), by a timer, or by both. The static physics-informed model falls from F1 0.768 without drift to 0.396 at severity 0.5 and 0.047 at severity 2. From severity 0.5 the best online variant has the higher point estimate, but the 95% intervals separate only from severity 1.5, where F1 is 0.313 against 0.144. Without drift the static model is ahead. Online adaptation therefore mitigates drift without curing it, because it needs timely labels and enough failure examples in each window. The study contributes a reproducible drift-aware evaluation protocol, an ablation of feature engineering, and a candid account of where adaptive learning helps and where it does not.

Keywords: predictive maintenance, concept drift, online learning, class imbalance, gradient boosting, industrial IoT, fault detection, AI4I 2020.