Ensuring the long-term reliability of critical digital instrumentation and control systems is paramount for the operational safety of nuclear power plants (NPPs). However, these digital components often exhibit cliff-edge characteristics where functional integrity may degrade without observable external precursors until an abrupt failure occurs at a specific threshold. Current periodic surveillance strategies primarily rely on discrete pass/fail criteria during scheduled tests, which inherently limits the ability to track gradual performance degradation or proactively mitigate potential risks between intervals. To bridge these monitoring gaps and the discrete nature of available diagnostic data, this paper proposes a deep learning-based remaining useful life (RUL) prediction framework that integrates accelerated aging experimental data with a novel dynamic data synthesis technique.
In this study, accelerated aging tests were conducted at 100°C on safety-grade digital input cards from the POSAFE-Q platform used in reactor protection systems to characterize threshold voltage drift driven by semiconductor degradation mechanisms, such as hot carrier injection and negative bias temperature instability. Based on the experimental results, multifaceted physical reference indicators were established for prognostic modeling, including a functional failure boundary defined at 10.2V and an early warning threshold derived from 500 cumulative damage events.
A significant bottleneck in developing data-driven RUL models for nuclear applications is the severe scarcity of failure data because acquiring representative dynamic environmental datasets would require decades of operational history. This paper overcomes this challenge by introducing a dynamic data synthesis approach that statistically approximates the complex operational profiles of NPPs, including seasonal temperature variations, load fluctuations, and 18-month periodic maintenance overhauls, by synthesizing the Arrhenius law with Miner's rule. Specifically, by generating 10,000 unit-specific time-series scenarios through Latin hypercube sampling to account for manufacturing variances and thermal uncertainties within instrumentation cabinets, the framework ensures robust generalization across diverse stress conditions. This synthesis process incorporates stochastic variations in thermal resistance and time constants, enabling the deep learning architecture to capture the synergistic effects of cumulative thermal stressors.
The proposed framework utilizes advanced deep learning architectures to learn non-linear temporal degradation patterns and assists risk-informed decision-making by providing quantitative uncertainty bounds for its predictions. Experimental evaluations demonstrate that the model achieves high RUL prediction accuracy across heterogeneous operational scenarios, confirming its capacity to visualize latent degradation states decades before functional failure. This research provides a reliable diagnostic methodology to facilitate the transition from time-based maintenance to condition-based maintenance in the nuclear industry while simultaneously alleviating the combinatorial burden of long-term data acquisition.