Calibration study of data-driven fault diagnosis models for complex industrial systems: A case study of nuclear power plant

Although deep learning has emerged as a prominent paradigm for data-driven fault diagnosis, offering high efficiency and strong diagnostic capability, its deployment in safety-critical industrial applications is still limited by the lack of reliable model calibration. Specifically, data-driven fault diagnosis models often exhibit pervasive overconfidence when handling high-dimensional and strong-nonlinear industrial process data, which can mislead operators into erroneous actions and undermine operational safety. To address gaps in existing research, a systematic investigation is conducted for the first time into both miscalibration mechanisms and the applicability of calibration strategies under realistic practical challenges, including sample imbalance, domain drift, and unknown faults scenarios. Using a representative nuclear power plant diagnosis setting as the test case, the interplay between model configurations and calibration performance is elucidated, revealing that modification of the model architecture alone is insufficient to reliably alleviate miscalibration and that dedicated calibration strategies remain necessary. Furthermore, representative calibration methods based on different principles are comparatively evaluated under challenging operating conditions, and both their effectiveness and limitations are analyzed. The results provide scenario-grounded evidence on how calibration methods behave in complex, high-dimensional, and strong-nonlinear industrial fault diagnosis tasks, while also clarifying their applicability and limitations in the investigated setting. Overall, this study offers a more cautious and practically relevant understanding of model calibration for trustworthy fault diagnosis in safety-critical industrial systems.