Predictive Maintenance of Industrial Pumps in Manufacturing Systems Using Hybrid Wiener–Hammerstein Ensemble Model

Centrifugal pumps are essential assets in industrial installation, where they support continuous dynamic operation in manufacturing processes, power plant, water distribution system facilities. In a small and medium scale industries, pump failure/faults are often detected only after noticeable performance deterioration, resulting in increased unplanned down time, energy consumption and higher operational risks. Traditional fault predictive maintenance strategies typically rely on single-sensor vibration analysis or standalone statistical indicators, which tend to lose reliability when nonlinear behaviour arises from coupled hydraulic and electrical interactions. To overcome this limitations, this study proposes a Hybrid Wiener–Hammerstein ensemble method for predictive maintenance (HWE–PM) that integrate physical system dynamics with data driven learning approach. The primary premise is that incipient faults are reflected not only in conventional signal features but also in structured residual patterns that signify deviation from normal operating conditions. A nonlinear-linear-Wiener-Hammerstein representation is employed to organize time and frequency domain information while an ensemble of heterogeneous classifiers–Random Forest (RF), XGBoost andk-Nearest Neighbours (kNN) is used to learn fault specific decision boundaries. The probabilistic output of these base learners is subsequently combined through a logistic regression meta learner using cross validated stacking, maximizing prediction stability under various operating conditions. The proposed work is evaluated using synthetic multimodal dataset developed to emulate the operating characteristics of a Crompton Greaves CG-IPM-15, three phase centrifugal pump and instrumented with 26 sensing channels. The HWE-PM model achieves an overall classification accuracy of 97.6% with a low false alarm rate of 2.3%, outperforming all individual base classifiers. The results demonstrate effective discrimination among multiple degradation modes, including bearing wear, aging of insulation, impeller imbalance and cavitation. Further implementation and model deployment on Jetson Nano and ESP32 edge devices confirms the feasibility of real time fault monitoring environment. Overall, the proposed HWE-PM based framework offers an interpretable and scalable predictive maintenance solution aligned with the requirement of Industry 4.0 environments.