Session: 01-04: Nuclear Plant Operation, Modification, Life Extension, Maintenance and Life Cycle - IV

Paper Number: 134882

134882 - Attention Mechanisms Based Advancing Interpretable Machine Learning Method for Nuclear Power Plant Fault Diagnosis 

Abstract: 

In the critical domain of nuclear power plant operations, ensuring safety and operational efficiency is paramount, and it primarily hinges on effective and reliable fault diagnosis. Traditional machine learning models, such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs), have demonstrated considerable potential in fault diagnosis due to their high computational power and pattern recognition capabilities. However, these models often suffer from a 'black-box' nature. This characteristic limits their practical applicability in nuclear power fault diagnosis due to a lack of transparency. The absence of interpretability in these models poses significant challenges, as it's crucial for operators to understand the reasoning behind the model's predictions to ensure safety and compliance with regulatory standards.

 

Addressing this challenge, our study introduces a novel approach by applying Shapley value theory from game theory to analyze the reasoning process of fault diagnosis models. The Shapley value, a concept from cooperative game theory, provides a fair distribution of rewards among players based on their individual contributions to the total payout. In the context of machine learning, it offers a method for attributing the prediction of a model to its input features, thereby demystifying the decision-making process of complex models.

 

Our research begins with the establishment of a convolutional neural network as a baseline fault diagnosis model applied to nuclear power. This CNN model is trained and validated on a comprehensive dataset comprising various fault scenarios in nuclear power plants. Through the utilization of the SHapley Additive exPlanations (SHAP) method, we conduct interpretability analysis for each fault diagnosis case. The SHAP method offers a unified measure of feature importance, providing insights into how each feature contributes to the model's prediction. The interpretability analysis results are visually displayed in this study, offering intuitive and understandable explanations of the model's decisions.

 

By analyzing the results obtained from the SHAP method, we can discern the contribution of each feature to the final result in different cases. This analysis reveals critical insights into the model's decision-making process, highlighting the features that play a pivotal role in fault detection and diagnosis.

 

Building upon these insights, our study further enhances the baseline model by integrating an attention mechanism. This mechanism, inspired by the human cognitive process, allows the model to focus on the most relevant features for making predictions. We propose an innovative and interpretable nuclear power fault diagnosis model that combines the strengths of CNNs with the enhanced focus provided by the attention mechanism. The attention mechanism learns and optimizes the weight distribution of the neural network by combining the results of interpretability analysis. This integration not only enhances the model's prediction accuracy but also its interpretability, making it more suitable for practical applications in nuclear power plants.

 

Our approach was rigorously tested through a series of experiments utilizing simulation data representative of various fault conditions in nuclear power plants. The model's performance was benchmarked against established baseline methods, demonstrating its effectiveness and efficiency. The results of our experiments reveal that our proposed method not only significantly improves the performance of the nuclear power fault diagnosis model but also offers enhanced interpretability. This aspect is particularly important for operators and engineers who rely on these models for critical decision-making processes.

 

In conclusion, this research contributes a vital reference for employing machine learning methods in the field of nuclear power fault diagnosis. It potentially transforms practices by combining advanced analytical techniques with interpretability and reliability. Our proposed model stands as a significant advancement in the field, offering a promising avenue for future research and application in nuclear power plant operations and safety management.

Presenting Author: Jie Liu Technical University of Munich

Presenting Author Biography: PhD candidate at the Chair of Nuclear Technology, Technical University of Munich, engaged in research on nuclear power fault diagnosis technology.

Authors:

Jie Liu Technical University of Munich
Rafael Macián-Juan Technical University of Munich