<p>The adoption of Maintenance 5.0 signifies a shift towards an advanced level of human-centered asset management that prioritizes self-sufficiency, resilience, and sustainable practices. This study introduces the attention-enhanced deep learning expert system that combines spatiotemporal convolutional neural networks (CNNs) and gated recurrent units (GRUs) for predictive maintenance (PdM) in mission-critical oil and gas assets-uniquely aligned with the principles of Maintenance 5.0 and validated across simulation, benchmark, and representative field datasets. The architecture processes multi-modal sensor data, including vibration, acoustic, thermal, and operational signals, while an attention mechanism is incorporated to improve interpretability and feature relevance. The model was validated on three datasets: the Tennessee Eastman Process (TEP) simulation, the Vishwakarma Institute of Technology (VIT) rolling element bearing vibration dataset, and a simulation‑derived offshore oil‑and‑gas field gas‑compressor operations dataset. Results demonstrate that the system detects faults with high accuracy, high sensitivity and greater computational efficiency than conventional approaches and baseline deep models. Specifically, it achieved F1 scores of 96.85% and 87.0% (with accuracies of 97.45% and 92.8%) on the bearing and TEP datasets, respectively. By addressing key challenges of interpretability, model generalization, and scalability, the system offers a viable Al-driven PdM solution. It supports sustainable asset lifecycle management and can considerably reduce unscheduled downtime and operating expenses (OPEX). As industry transitions towards Industry 5.0, this work provides a versatile and forward-looking framework for intelligent Condition-based Maintenance (CBM) in resource-intensive, safety-critical environments such as the oil and gas sector, with potential for extension to other industrial domains.</p>