An extreme value theory-based generative distribution calibration framework for open-set fault diagnosis in thermal engineering systems

Fault diagnosis in thermal engineering systems is inherently an open-set problem, since unforeseen fault modes may emerge during operation and are not available during model training. Conventional closed-set methods therefore tend to misclassify unknown faults. Existing open-set recognition approaches, such as OpenMax, Hypersphere-based methods and DeepSVDD, still exhibit limited unknown-fault discrimination, noise robustness, and adaptability to complex operating conditions. which can pose challenges to safe operation. To address these issues, the EVT-GDC (Extreme Value Theory-based Generative Distribution Calibration) framework for OSFD (Open-Set Fault Diagnosis) in thermal systems is proposed. EVT-GDC employs a three-stage learning strategy. Firstly, a CNN (Convolutional Neural Network) is trained to learn discriminative fault features. Subsequently, a lightweight MLP (Multi-Layer Perceptron)-based class-conditioned generator is introduced to model class-wise feature distributions and generate category-consistent synthetic monitoring samples. Finally, generated and real samples are jointly utilized to calibrate feature distributions, resulting in more compact intra-class representations and clearer separation between known and unknown fault regions. This calibrated feature space is further combined with EVT-based open-set modeling for unknown fault identification. The proposed method is evaluated on the AHIAVF (Advanced Hua-long Pressurized Reactor Intelligent Algorithm Validation Facility) dataset, a high-dimensional and highly nonlinear benchmark for thermal engineering systems. EVT-GDC consistently outperforms representative methods, including OpenMax,Hypersphere-based methods and DeepSVDD. Under low-openness scenarios, the proposed method achieves F1-score improvements of 7.95%-14.82%. Even under high-openness scenarios, it maintains improvements of 4.76%-9.95% and showed improved noise robustness in the tested scenarios. These results indicate the potential of EVT-GDC for OSFD in complex thermal systems.