Data-Driven Multivariate Model for condition monitoring of 100 MW Steam Turbines

Abstract

A data-driven diagnostic model was developed for 100 MW steam turbines, integrating Principal Component Analysis (PCA) and Multivariate Statistical Process Control (MSPC) to characterize the normal functional-dynamic variability of the system and to detect early operational deviations. Historical databases from the online monitoring system, containing vibration and technological measurements acquired over more than two years of continuous operation, were used. The methodological process included data preprocessing, conditioning, cleaning, and validation, followed by the construction of the multivariate model and the definition of a reference pattern sample. Results demonstrated the statistical stability of the model and its ability to discriminate abnormal conditions using Hotelling’s T² and Q-residuals statistics. The proposed approach improved early fault detection and contributed to the implementation of condition-based predictive maintenance strategies, providing an effective tool for functional diagnostics of high-power industrial turbines and establishing a methodological framework that can be extended to other rotating thermal systems.