Gas turbine gas path fault diagnosis based on open-set recognition and physics-data fusion

Condition-Based Maintenance relies on accurate diagnosis of gas-path faults. Existing data-driven methds (DDMs) are limited by insufficient training classes. Model-based methods (MBMs) struggle with diagnostic accuracy due to insufficient model fidelity. This study presents a novel hybrid diagnostic framework that combines open-set recognition (OSR) with model-based diagnosis to address existing limitations. As the first application of OSR in gas path fault diagnosis, this approach employs a class-conditional autoencoder to classify known fault classes while also identifying unknown classes. A complementary diagnostic pathway, driven by physical models and expert knowledge, is also developed. This pathway incorporates the Unscented Kalman Filtering, Cumulative Sum, and a Fuzzy Inference System to classify fault classes without relying on historical data. Building on this, a physics-data decision fusion strategy is proposed, which combines the outputs of both paradigms using evidential reasoning rules. This fusion approach evaluates the weight of evidence by considering the information entropy of each diagnostic path's prediction results and assesses the reliability of the evidence based on the unknown class probability and classification accuracy, respectively. By integrating both the weight and reliability of evidence, the fusion strategy prioritizes the diagnostic path with higher accuracy for known classes, while lever-aging the more generalizable model-based diagnostic results for unknown classes, thereby enhancing overall diagnostic performance. Comprehensive experimental validation demonstrates that the proposed framework has superior fault diagnosis ability compared to conventional methods. This physics-data fusion methodology provides an effective solution for actionable gas-path fault diagnosis in real-world operational scenarios, where fault samples may be incomplete.