Unsupervised Anomaly Detection in Industrial Gas Turbines

Sammanfattning

The transition to renewable energy sources demands increased operational flexibility from industrial gas turbines, which often operate under varying load conditions to stabilize the grid. This operational variability renders traditional, static anomaly detection thresholds ineffective, as a global model of normality fails to account for the inherent variance between quiet steady-state operations and noisy transient phases. This thesis addresses the critical challenge of detecting acoustic anomalies in gas turbines operating under such non-stationary conditions by proposing a novel, unsupervised, regime-specific deep learning framework, in which each regime corresponds to a distinct turbine operating state. The core contribution of this work is an unsupervised “Regime Router” that utilizes a Gaussian Mixture Model (GMM) to automatically partition the complex operating envelope into distinct regimes without requiring manual labels. This router directs acoustic data to a mixture-of-experts architecture where separate Convolutional Autoencoders (CAEs) are trained to model the specific acoustic manifold of each regime. The proposed framework was evaluated on a large-scale, real-world dataset from Siemens Energy's SGT-700 turbines, comprising over 280,000 spectrogram windows of normal operation and 60,000 windows of unseen anomalous behavior. The results demonstrate that the regime-specific strategy significantly outperforms a global baseline in terms of operational viability. While the global model achieved high aggregate metrics, it failed in practice due to a single, incompatible decision threshold that was too loose for quiet regimes (missing subtle faults) and too tight for noisy transients (causing false alarms). By applying adaptive, regime-specific thresholds, the proposed system achieved near-perfect separability (ROC-AUC ≈ 1.000) in steady-state regimes while effectively managing the noise floor during transient startup phases. Furthermore, the study confirmed that the Structural Similarity Index (SSIM) loss function provides superior sensitivity to acoustic structural anomalies compared to standard pixel-wise reconstruction errors such as Mean Absolute Error (MAE) and Mean Squared Error (MSE). This work establishes a validated pathway for context-aware, unsupervised predictive maintenance in flexible power generation systems.