Enhancing Nuclear Power Plant Monitoring through Deep Learning-Based Object Detection

Abstract

This study delves into the potential of employing deep learning for object detection as a means of monitoring abnormal states in components of nuclear power plants. Experiments were carried out using a scaled-down nuclear power plant experimental facility (URI-LO) equipped with a drone capturing visual and thermal images. Abnormal conditions were simulated by manipulating water levels in the steam generator and stopping fluid circulation pumps. A comparative analysis of various deep learning models was conducted to identify the optimal model for object detection. The results demonstrated that the YOLOv8m model exhibited superior accuracy in detecting anomalies. This research contributes to the development of advanced monitoring systems for nuclear power plants, leveraging deep learning techniques for real-time anomaly detection.

Introduction

Nuclear power plants, with their intricate systems and components, demand continuous monitoring to ensure safety and reliability. Conventional monitoring methods often rely on manual inspection and analysis, which can be time-consuming and prone to human error. Advanced techniques that harness artificial intelligence, particularly deep learning, have emerged as promising tools for enhancing monitoring capabilities. This study aims to explore the application of deep learning for object detection in nuclear power plant monitoring, with a focus on identifying abnormal states in critical components.

Experimental Setup

A scaled-down nuclear power plant experimental facility (URI-LO) served as a platform to simulate real-world conditions. This facility comprised key components of a nuclear power plant, including a reactor pressure vessel, steam generators, reactor coolant pumps, and various pipelines. A drone equipped with a vision camera and an infrared (IR) camera was employed to capture visual and thermal images of the facility.

Abnormal State Experiments

To simulate abnormal operating conditions, two types of experiments were conducted. In the first experiment, the water level in the steam generator was gradually decreased to simulate a scenario where heat removal is compromised. In the second experiment, the fluid circulation pumps were stopped to disrupt heat transfer. The drone captured images of the facility during both normal and abnormal states.

Deep Learning for Object Detection

Deep learning for object detection was employed to analyze the captured images. The goal was to develop a model capable of identifying abnormal states based on visual and thermal cues. Various deep learning models, including one-stage (YOLOv5, YOLOv8, NanoDet) and two-stage (Mask R-CNN) models, were trained and evaluated.

Model Evaluation

The performance of the deep learning models was assessed using metrics such as mean average precision (mAP) and floating-point operations per second (FLOPS). The YOLOv8m model demonstrated superior performance, achieving the highest mAP score and exhibiting efficient inference time.

Results and Discussion

Analysis of the experimental results revealed that the YOLOv8m model effectively detected abnormal states in the nuclear power plant components. The model accurately identified changes in water level in the steam generator and the stoppage of fluid circulation pumps. This demonstrates the potential of deep learning for object detection in monitoring nuclear power plant components in real time.

Conclusion

The study demonstrates the feasibility of utilizing deep learning for object detection to enhance monitoring capabilities in nuclear power plants. The YOLOv8m model showed promising results in detecting abnormal states in critical components, highlighting the potential for real-time monitoring and anomaly detection. This research contributes to the development of advanced monitoring systems for nuclear power plants, leveraging deep learning techniques to improve safety and reliability.

Keywords:

Nuclear power plant monitoring, Deep learning, Object detection, YOLOv8, Abnormal state detection, Thermal imaging.