AI-Powered Fault Detection in Gas Turbine Engines: Enhancing Predictive Maintenance in the U.S. Energy Sector

Abstract

In the dynamic energy sphere in the U.S., operational dependability of gas turbine engines becomes vital to the continuous power production and sustenance of the national grid. Peak and baseload power production, the mainstay of gas turbines, is exposed to high thermal, mechanical, and chemical stresses that lead to wear out of components over time. Traditionally, maintenance practices have either been reactive or time-based, based upon fixed intervals, which has led to premature replacement of components, or vice versa, undetected degradation leading to catastrophic failure. Such approaches carry high operating costs, lower turbine availability, and jeopardize plant safety. The central objective of this research is to build and test an AI-driven fault detection system to identify early signs of failure in gas turbine engines and apply it specifically to deployment within the energy sector of the U.S. Improving gas turbine performance is essential to raising the cost-effectiveness and sustainability of power production systems. The dataset used in this research entails high-resolution operational parameters gathered from different industrial gas turbines operating within U.S. energy facilities. Engine operational data includes multiple time-dependent measurements that monitor essential parameters like turbine temperature at the inlet and outlet and rotational speed, torque measurements alongside vibration levels and power output, fuel rate and intake pressure, and exhaust temperature and oil temperature. Real-time data acquisition from embedded sensor arrays allowed researchers to track turbine performance throughout changing operational states at sub-minute data points. A centralized time-stamping system maintains channel synchronization, thus allowing analysts to draw accurate conclusions about operational states throughout the recorded period. To support strong and interpretable fault detection, we utilized a variety of machine learning models, each chosen for its specific strength at discriminating operations from fault conditions in gas turbines. We used a multi-metric-based evaluation approach that combined statistical validity with operational applicability to assess each model's fault detection capability. Logistic Regression attained the highest accuracy, followed very closely by Random Forest. XG-Boost attained the lowest accuracy of all three algorithms. The use of AI-driven fault detection under predictive maintenance has the potential to revolutionize U.S. power plants using gas turbines as the primary source of electricity generation. In the fiercely competitive and heavily regulated environment of the U.S. energy industry, fault anticipation presents the key to competitive advantage. Using AI-based diagnostics reduces manual checks and simplifies the servicing process by prioritizing technician resources to confirmed at-risk components. AI-powered fault detection is critical to improving grid resilience within the U.S. energy infrastructure by assuring that peak-demand-balancing and grid-stabilizing gas turbines remain fault-free.