Natural Noise Feature Extraction in Sensor Data for Enabling Digital Twins and Real-Time Anomaly Detection

Abstract The automation of turbomachinery control systems and the increasing adoption of data-driven solutions in their design, operation, and maintenance have transformed these systems into complex cyber-physical entities. While this evolution enhances functionality and efficiency, it also introduces new challenges, particularly in anomaly detection and system reliability. Machine learning, a widely employed method for anomaly detection, relies on extensive datasets representing both normal operations and faulty conditions. However, collecting such data is often expensive, time-consuming, and impractical, especially when destructive testing is involved. In this context, Synthetic Data Generation (SDG) emerges as a promising solution to address data scarcity by augmenting real-world datasets for training machine learning algorithms. Despite its potential, the application of realistic synthetic data generation (R-SDG) in turbomachinery remains limited due to challenges such as the inherent noise in sensor data and the noise-free nature of simulation data. This study addresses these limitations by focusing on the extraction of natural noise characteristics in turbomachinery. This effort contributes a critical step toward advancing R-SDG by designing and testing a framework for noise feature extraction. Furthermore, the article provides an extracted natural noise model that can be used to inject realistic noise into synthetically generated data (via numerical or physical models). The findings from the case study on gas turbines in electricity generation systems demonstrate the potential of this approach to facilitate data augmentation for ML training, support digital twin development, and enable real-time anomaly detection in similar turbomachinery systems.