In nuclear power plants (NPPs), the reliability of sensor signals is important for operators' situational awareness and for ensuring safe operation. Operators make decisions based on information collected from various instrumentation sensors, which serve as inputs for artificial intelligence (AI)-based operator support systems. However, signal faults caused by sensor malfunctions, aging, and environmental factors can occur in actual operating environments. These faults may delay accident recognition or cause misdiagnosis, increasing human error risk. Signal integrity is particularly important in emergency situations, where rapid decision-making is imperative. This study proposes an AI-based algorithm for effective identification and restoration of sensor signal faults during emergencies in NPPs. First, the algorithm verifies the input signals to detect faults. Subsequently, it selectively restores only faulty signals. The restored signals are then used for accident diagnosis, preventing performance degradation caused by faulty inputs. The algorithm was evaluated using artificially generated data for three types of faults: bias, drift, and stuck. Results demonstrated high accuracy in fault detection and restoration. Additionally, restored signals enabled accurate accident classification. This study is expected to enhance NPP safety and reliability by mitigating the impact of signal faults on AI-based operator support systems and decision-making.

The reactor protection system (RPS), which is an integral component of the safety system in nuclear power plants, can pose a substantial risk to plant safety in the event of an unexpected failure. Although RPS integrity is currently evaluated via self-diagnostic functions, periodic tests, and overhaul procedures, these methods are limited in scope and are difficult to verify during normal operation. Therefore, this study proposes a deep learning-based approach for predicting potential RPS failures, even during normal operation. Accordingly, critical electronic components within the RPS were selected with a particular focus on predicting the remaining useful life of the photocoupler. A deep-learning-based prediction model was developed using data obtained from accelerated aging tests in conjunction with temperature change scenarios generated to reflect actual operating conditions. Specifically, a long short-term memory network integrated with Monte Carlo dropout was applied to estimate the remaining useful life and quantify the prediction uncertainty. In addition, the performance of the model was enhanced by incorporating physics-informed regularization into the loss function. The effectiveness and applicability of the proposed model were verified via performance comparisons with alternate models that utilize different deep learning methods and loss functions.

A reactor protection system (RPS) is a core safety system for the stable operation of nuclear power plants (NPPs). Modern RPSs adopt digital platforms based on programmable logic controllers, offering enhanced reliability and maintainability compared with analog systems. Despite these advancements, current practices such as self-diagnosis functions, periodic operational tests, and scheduled maintenance remain limited in assessing component conditions during normal operation. To address this limitation, this study proposes an artificial intelligence (AI)-based condition-monitoring framework to improve the maintenance strategy for the internal electronic components of digital RPSs. The framework combines a rule-based model to ensure data integrity with an AI-based model to monitor component condition. The methodology was specifically applied to the photocoupler, a critical electronic component whose failure can significantly affect the digital RPS. To train and validate the AI-based model, a scenario dataset was generated using accelerated aging data for photocouplers to simulate the actual operating environment. In addition, a conceptual monitoring interface was designed to evaluate the practical applicability of this approach. The framework early detected the faults of photocouplers in RPS and accurately diagnosed their conditions. It is expected that the proposed framework can enable condition-based maintenance, reduce unnecessary inspections, and improve system availability.

Although the probability of a severe accident at a nuclear power plant is low, it can result in catastrophic outcomes. This study employed deep-learning-based time-series models to simultaneously predict the core exit temperature, containment pressure, and hydrogen concentration, which are critical monitoring variables during severe accidents. Four models (recurrent neural network, long short-term memory, convolutional neural network (CNN), and temporal convolutional network) were implemented in a multi-input multi-output structure and trained on simulation data from cold-leg loss-of-coolant accident (LOCA), hot-leg LOCA, and steam generator tube rupture scenarios. To address predictive uncertainty, Monte Carlo dropout was applied to estimate the confidence intervals. Among the models, the CNN demonstrated a superior balance between predictive accuracy and computational efficiency. It achieved highly competitive performance, despite having significantly fewer trainable parameters and a dramatically faster training time. This approach combines multivariate prediction and uncertainty quantification, demonstrating the practical potential for integration into AI-based operator support systems. This methodology is expected to enhance the situational assessments of operators and support proactive mitigation strategies. Future work will expand the scope of validation by incorporating a wider range of accident scenarios and operational conditions, while also accounting for external and environmental variables that may influence the prediction accuracy.

Explainable artificial intelligence (XAI) is employed to clarify the rationale behind AI outputs and resolve black-box nature of artificial intelligence (AI). This is intended to enhance trustworthiness and usability of AI-based technology for operation or decision-making support. However, XAI explanations now are often more accessible to developers, who construct, verify, and optimize AI models, than to operators, who need to understand and employ these models for decision-making. Therefore, this study aims to develop a Grad-CAM-based deep learning methodology that provides operator-centered explanations for enhancing explainability of AI outputs and the trustworthiness of the AI technologies for operation support in the context of performing procedure-based operating tasks. In this study, an XAI model based on gradient-weighted class activation mapping and a dilated causal convolutional neural network was developed to identify abnormal states and provide operator-centered explanations within the scope of abnormal operating procedures corresponding to identified abnormal states. Furthermore, representation of operator-centered explanations was addressed to effectively display the AI-supported information from the proposed model on human-system interface for operating tasks.

Severe accidents in nuclear power plants progress in complex and non-linear ways, making it difficult for operators to make quick and accurate situational assessment. To support proactive operator decision-making, this study constructed a deep learning model to predict the future behavior of safety variables in multi-steps during a severe accident. Based on the patch time series Transformer model, which shows high performance in time-series prediction, various severe accident scenario data generated by the modular accident analysis program simulation code were used for training. To improve the model’s performance and robustness, a complex noise injection technique simulating instrument uncertainty was applied. Furthermore, the prediction uncertainty was quantified using Monte Carlo dropout. Performance evaluation results showed that the proposed model effectively predicted the dynamic behavior of the safety variables. In particular, the model trained with injected noise showed significantly superior robustness on noisy test data compared to the model trained on clean data, suggesting high applicability in real operational environments. In conclusion, the prediction framework proposed in this study confirmed its potential to be utilized as part of an operator support system that provides reliable future information as a decision support tool for operators during a severe accident.

A Reactor Protection System (RPS) is composed of numerous electronic components and requires a high level of reliability. A failure in photocouplers, among these components, can compromise the isolation function of the monitoring system and control circuits in the RPS, posing a serious threat to its reliability. Therefore, maintaining the integrity of photocouplers and complementing conventional maintenance practices by enabling proactive replacement before failure occurs is of great importance. To address this issue, this study proposes an AI based component condition diagnosis approach. Data collected from accelerated aging tests of photocouplers were utilized for this purpose and converted to reflect actual environmental conditions. Furthermore, based on the InceptionTime architecture specialized for time series classification, an improved model structure was developed by integrating a Squeeze-and-Excitation (SE) block after each Inception module to reflect inter-channel importance. Using this enhanced structure, the model classified the component states into three categories; normal, monitoring, and replacement-recommended states. Additionally, the t-distributed Stochastic Neighbor Embedding (t-SNE) technique was employed to visualize the distribution of state dates within the learned latent space, thereby qualitatively confirming that the model effectively distinguishes features among different states. Moreover, by visualizing the channel weights of the SE block, the relative importance of each input variable was analyzed, revealing that the model assigns higher weights to key degradation related features. These results demonstrate that the proposed approach is suitable for diagnosing the condition of RPS components and securing explainability in AI based diagnostic systems.

In nuclear power plants, various alarm systems are activated in the event of abnormal or emergency situations to notify operators of the abnormal situation. However, most of these alarm procedures initiate operator situation awareness and response only after the abnormal situation has already occurred. This can present challenges for operators in quickly recognizing and responding to the situation. Early recognition of state changes and making appropriate judgments during abnormal situations are crucial for ensuring the safety of nuclear power plants. However, there are limitations to relying solely on the cognitive abilities and judgment of operators for a swift response. To address this, many studies are being conducted to utilize artificial intelligence to assist operators. This study proposes a proactive alarm prediction technology that can overcome transient situations by reducing the operator’s workload and supporting decision-making through preemptive problem recognition in real-time transient conditions. To achieve this, a variational auto-encoder is used to detect subtle, early signs of anomalies in the plant. Furthermore, a transformer technique predicts the timing of alarm occurrences, providing this information to the operators. This is expected to enhance the situational awareness of operators, prevent human error, and ultimately contribute to improving the safety of nuclear power plants.

The entry condition of the Severe Accident Management Guideline (SAMG) in Nuclear Power Plants (NPPs) is determined by the Core Exit Temperature (CET). If the CET exceeds 922 $K$ (1200 ${}^{\circ}F$ ), severe accident management begins. Because a severe accident can induce a large scale of damage, it is necessary to prepare for such accident and take preemptive actions. However, the operators may be confused by the complexity of the system, which can delay actions. Therefore, operators need the entry time information of SAMG to act proactively. In this study, the entry time was predicted through CET prediction. The Explainable Boosting Machine (EBM) model was used to select the input variables and the Long Short-Term Memory (LSTM) model was used to predict CET 600 seconds ahead. And the Monte Carlo (MC) dropout method was used to evaluate the uncertainty of the predictions at a 95% confidence level. As a result, the LSTM model performed well and the evaluated uncertainty provided confidence in the predictions with a confidence interval. Predicting 600 seconds ahead provides time for the operators to take actions on the accident, and the uncertainty evaluation adds reliability to the model's prediction. The results of this study are expected to be used as part of the operator support system and serve as a means for rapid accident mitigation actions. Furthermore, the integration of AI-based predictive models and uncertainty evaluations ensures that operators are equipped with reliable information, enhancing their ability to act preemptively and effectively in response to severe accidents.

In nuclear power plants (NPPs), it is important to ensure the validity of signals for safe operation. However, signals can be corrupted by aging and environmental factors. Thus, active research in signal verification and restoration is required. Previous signal failure detection studies have typically treated any anomalous data as a signal failure and proceeded with restoration. Because these studies targeted only signal failure data, this approach was taken. However, true abnormal situations will not be recognized if unlearned actual abnormal data is treated as a signal failure and restored. Therefore, it is necessary to distinguish between signal failures and abnormal situations. In this study, an algorithm was proposed to distinguish between signal failures and abnormal situations. The algorithm was implemented using an autoencoder (AE) and a long short-term memory (LSTM)-AE. Data from the compact nuclear simulator (CNS) was used for training and testing. The results demonstrated that the proposed algorithm effectively distinguished between signal failures and abnormal situations. Additionally, the LSTM-AE performed better compared to the AE.

It is difficult to detect a small-scale leakage in a nuclear power plant (NPP) quickly and take appropriate action. Delaying these procedures can have adverse effects on NPPs. In this paper, we propose leak flow rate prediction using the bidirectional gated recurrent unit (Bi-GRU) method to detect leakage quickly and accurately in small-scale leakage situations because large-scale leak rates are known to be predicted accurately. The data were acquired by simulating small loss-of-coolant accidents (LOCA) or small-scale leakage situations using the modular accident analysis program (MAAP) code. In addition, to improve prediction performance, data were collected by distinguishing the break sizes in more detail. In addition, the prediction accuracy was improved by performing both LOCA diagnosis and leak flow rate prediction in small LOCA situations. The prediction model developed using the Bi-GRU showed a superior prediction performance compared with other artificial intelligence methods. Accordingly, the accurate and effective prediction model for small-scale leakage situations proposed herein is expected to support operators in decision-making and taking actions.

In abnormal states of nuclear power plants (NPPs), operators undertake mitigation actions to restore a normal state and prevent reactor trips. However, in abnormal states, the NPP condition fluctuates rapidly, which can lead to human error. If human error occurs, the condition of an NPP can deteriorate, leading to reactor trips. Sudden shutdowns, such as reactor trips, can result in the failure of numerous NPP facilities and economic losses. This study develops a remaining trip time (RTT) prediction system as part of an operator support system to reduce possible human errors and improve the safety of NPPs. The RTT prediction system consists of an algorithm that utilizes artificial intelligence (AI) and explainable AI (XAI) methods, such as autoencoders, light gradient-boosting machines, and Shapley additive explanations. AI methods provide diagnostic information about the abnormal states that occur and predict the remaining time until a reactor trip occurs. The XAI method improves the reliability of AI by providing a rationale for RTT prediction results and information on the main variables of the status of NPPs. The RTT prediction system includes an interface that can effectively provide the results of the system.

Welding processes are used to connect several components in nuclear power plants. These welding processes can induce residual stress in welding joints, which has been identified as a significant factor in primary water stress corrosion cracking. Consequently, the assessment of welding residual stress plays a crucial role in determining the structural integrity of welded joints. In this study, a deep fuzzy neural networks (DFNN) with a rule-dropout method, which is an artificial intelligence (AI) method, was used to predict the residual stress of dissimilar metal welding. ABAQUS, a finite element analysis program, was used as the data collection tool to develop the AI model, and 6,300 data instances were collected under 150 analysis conditions. A rule-dropout method and genetic algorithm were used to optimize the estimation performance of the DFNN model. DFNN with the rule-dropout model was compared to a deep neural network method, known as a general deep learning method, to evaluate the estimation performance of DFNN. In addition, a fuzzy neural network method and a cascaded support vector regression method conducted in previous studies were compared. Consequently, the estimation performance of the DFNN with the rule-dropout model was better than those of the comparison methods. The welding residual stress estimation results of this study are expected to contribute to the evaluation of the structural integrity of welded joints.

In nuclear power plants, reactor coolant leakage can occur due to various reasons. Early detection of leaks is crucial for maintaining the safety of nuclear power plants. Currently, a detection system is being developed in Korea to identify reactor coolant system (RCS) leakage of less than 0.5 gpm. Typically, RCS leaks are detected by monitoring temperature, humidity, and radioactivity in the containment, and a water level in the sump. However, detecting small leaks proves challenging because the resulting changes in the containment humidity and temperature, and the sump water level are minimal. To address these issues and improve leak detection speed, it is necessary to quantify the leaks and develop an artificial intelligence-based leak detection system. In this study, we employed bidirectional long short-term memory, which are types of neural networks used in artificial intelligence, to predict the relative humidity in the leakage area for leak quantification. Additionally, an optimization technique was implemented to reduce learning time and enhance prediction performance. Through evaluation of the developed artificial intelligence model's prediction accuracy, we expect it to be valuable for future leak detection systems by accurately predicting the relative humidity in a leakage area.

In nuclear power plants, coolant leakage occurs for various reasons. Leak detection is important to ensure safety of nuclear power plants. Currently, a detection system for an unidentified reactor coolant system(RCS) leakage of less than 0.5gpm is being developed in Korea. The RCS leakage is detected through changes in radioactivity, humidity, and temperature in the containment air, and water level of sump. For small leaks, the change in humidity and temperature due to water vapor is very small, making the leak very difficult to detect until the leak accumulates in the instrument.BR In order to solve these problems and increase the leak detection speed, it is necessary to develop a system capable of real-time detection using artificial intelligence. In this study, long short-term memory and bidirectional long short-term memory, which are types of recurrent neural networks among artificial intelligence methods, were applied to perform initial relative humidity prediction for leakage quantification. Also, an optimization technique that reduces learning time and improves prediction performance for the optimization of learning was applied. Finally, the prediction performance was evaluated using the developed model.

A nuclear power plant is a large facility composed of many components, and abnormal states occasionally occur in which components fail. In the event of abnormal states, if appropriate measures are not taken, the abnormal states can worsen and cause an unexpected reactor trip. Therefore, in order to provide the operator with key state information in case of abnormal states, abnormal diagnosis and trip variable prediction were performed based on multi-task learning (MTL). The MTL is a method of performing multiple tasks through a single model. Specifically, the progressive layered extraction method, one of the MTL structures, was applied. It efficiently transmits information between tasks through a gating network and progressive routing mechanism. The proposed model showed higher diagnosis accuracy and lower prediction error than the basic MTL model. If the key state information is provided to the operator through the proposed model, it will be able to contribute to reducing human error and preventing the aggravation of abnormal states.

When abnormal operating conditions occur in nuclear power plants, operators must identify the occurrence cause and implement the necessary mitigation measures. Accordingly, the operator must rapidly and accurately analyze the symptom requirements of more than 200 abnormal scenarios from the trends of many variables to perform diagnostic tasks and implement mitigation actions rapidly. However, the probability of human error increases owing to the characteristics of the diagnostic tasks performed by the operator. Researches regarding diagnostic tasks based on Artificial Intelligence (AI) have been conducted recently to reduce the likelihood of human errors; however, reliability issues due to the black box characteristics of AI have been pointed out. Hence, the application of eXplainable Artificial Intelligence (XAI), which can provide AI diagnostic evidence for operators, is considered. In conclusion, the XAI to solve the reliability problem of AI is included in the AI-based diagnostic algorithm. A reliable intelligent diagnostic assistant based on a merged diagnostic algorithm, in the form of an operator support system, is developed, and includes an interface to efficiently inform operators.