Abstract
Fault detection and isolation (FDI) in gas turbines is pivotal to maximizing efficiency, avoiding catastrophic damage, and minimizing unplanned outages. Yet data-driven FDI schemes remain brittle: they lack self-organizing, adaptable structures, rely on limited validation, and degrade under noise and uncertainty. We propose a Self-organizing Type-3 Fuzzy Rough Wavelet Neural Network (ST3FRWNN) that addresses these gaps. First, we introduce a bell-shaped Type-3 membership function that enhances uncertainty handling. Second, we train a neuro-fuzzy architecture with a hybrid Adam-Unscented Kalman Filter optimizer for fast convergence. Third, we embed a self-organizing mechanism that grows and prunes rules adaptively, yielding compact models without sacrificing accuracy. Evaluations on a high-fidelity 163-MW Siemens gas-turbine simulator (Case 1) show average FDI rates of 99.302% (detection) and 99.324% (isolation) and real acoustic-emission signals (Case 2). ST3FRWNN surpasses state-of-the-art fuzzy systems, including FSRE-AdaTSK and TSK-SRB, while using fewer rules, and remains competitive with deep learning baselines (Transformers, LSTMs, CNNs) at a fraction of their parameters. These results, together with robustness to 20-dB SNR and rigorous five-fold cross-validation, demonstrate superior computational efficiency and make ST3FRWNN a practical, deployable solution for real-world gas-turbine health monitoring.