Intelligent Predictive Maintenance System for Turbomachinery Networks via Edge AI

Thesis Abstract

This study addresses the rising efficiency and reliability challenges in turbomachinery networks by developing an intelligent predictive maintenance system that leverages edge AI for real-time health monitoring, fault detection, and prognosis under constrained industrial environments. The research aims to reduce unplanned downtime, extend asset life, and optimize maintenance scheduling through a decentralized processing architecture that minimizes latency and preserves data sovereignty. Specific objectives include (i) designing an edge-enabled data acquisition and processing pipeline for heterogeneous turbomachinery sensors, (ii) developing and validating machine learning models for remaining useful life (RUL) estimation and fault classification, (iii) integrating uncertainty quantification and sensor fault handling into the predictive framework, (iv) evaluating system performance against centralized cloud-based baselines, and (v) formulating an operational playbook for industrial deployment. The methodological approach combines a pragmatic mixed-methods design anchored in reliability engineering and data science. The population comprises five industrial turbomachinery networks across three refineries, with a target cumulative asset base of 120 machines and approximately 25 TB of historical sensor data. A stratified random sampling technique selects 40 representative units for longitudinal study, ensuring coverage of centrifugal compressors, axial turbines, and centrifugal pumps. Data collection employs edge-enabled health monitoring units that capture vibration, temperature, pressure, rotational speed, and energy consumption at 1 kHz sampling, supplemented by maintenance records and incident logs. Instruments include calibrated vibration sensors, high-resolution thermistors, and a standardized maintenance event diary, with data validation procedures aligned to ISO 13374 standards. The predictive models deploy a hybrid approach convolutional neural networks (CNN) and long short-term memory networks (LSTM) for feature extraction and temporal sequence modeling, Gaussian process regression for RUL uncertainty bounds, and gradient boosting for fault diagnosis. Model calibration uses a nested cross-validation scheme, and feature importance is assessed via SHAP (SHapley Additive exPlanations) values. Anomaly detection integrates one-class SVM and autoencoder techniques to identify sensor degradations and incipient faults. Ethical considerations address data privacy, cybersecurity of edge nodes, and operator safety. Data analysis proceeds in three layers (i) descriptive analytics to characterize sensor performance and maintenance history, (ii) predictive analytics to quantify RUL and fault probabilities, and (iii) prescriptive analytics to optimize maintenance scheduling under budget and workforce constraints. Statistical validation includes time-series cross-validation, ROC-AUC metrics for fault detection, root mean square error (RMSE) for RUL predictions, and calibration plots for probabilistic estimates. A comparative performance assessment against cloud-only benchmarks evaluates latency, bandwidth utilization, and total cost of ownership. The study also investigates the robustness of edge AI under intermittent connectivity through scenario simulations and fault injection experiments. The anticipated findings indicate that the edge-driven predictive maintenance framework delivers up to 25% improvement in maintenance effectiveness, a 15–20% reduction in unplanned downtime, and more accurate RUL estimations with 95% credible intervals that reliably bound uncertainty, while preserving data confidentiality and reducing cloud dependency. The contribution to knowledge lies in (a) a rigorously validated, deployable edge AI architecture for turbomachinery health management, (b) a performance-economics model for edge-based predictive maintenance in industrial networks, and (c) methodological insights into integrating multi-modal sensor data, uncertainty quantification, and interpretable AI in a constraint-aware maintenance setting. The study culminates in a deployment-ready framework coupled with an operational guideline that addresses integration with existing maintenance information systems, cybersecurity protocols for edge devices, and workforce training programs. The main conclusion is that edge AI-enabled predictive maintenance for turbomachinery networks can achieve meaningful reliability and efficiency gains while meeting practical considerations of latency, data governance, and operational risk. Recommendations include extending the model to accommodate sudden policy changes in maintenance scheduling, implementing continuous online learning with drift detection, and conducting longitudinal impact studies across additional asset classes to generalize the framework’s applicability.