A physics-informed multi-modal transformer framework for predictive maintenance and remaining useful life estimation of electrical submersible pumps.

Electrical Submersible Pumps (ESPs) are widely used in oil and gas production but are highly vulnerable to mechanical, electrical, hydraulic, and thermal degradation under harsh downhole conditions. Unexpected ESP failures can result in severe production losses, costly interventions, and reduced operational reliability, highlighting the importance of predictive maintenance and Remaining Useful Life (RUL) estimation. Existing prognostic approaches often rely on single-modality data and lack physical consistency in degradation modeling. This study proposes HF-ESPNet, a physics-informed multi-modal transformer framework for joint ESP failure prediction and RUL estimation. The proposed model integrates heterogeneous operational data, including electrical telemetry, pressure, temperature, vibration indicators, well metadata, and maintenance history within a unified transformer-based architecture. Domain-specific physics-informed constraints related to hydraulic behavior, thermal dynamics, and vibration degradation are incorporated into the learning objective to improve physical consistency and prognostic reliability. The framework was evaluated using an anonymized real-world ESP dataset containing more than 58,000 multivariate time-series samples under diverse operational conditions. Experimental results demonstrate that HF-ESPNet outperforms conventional machine learning, deep learning, and transformer-based baseline models in both failure prediction and RUL estimation tasks. Ablation analysis further confirms the importance of vibration-derived features and physics-informed constraints in improving predictive robustness and degradation modeling accuracy. The results demonstrate that combining multi-modal representation learning, transformer-based temporal modeling, and physics-informed learning provides a robust and interpretable framework for ESP prognostics, supporting improved maintenance scheduling, reduced unplanned downtime, and enhanced operational reliability in oil and gas production systems.