Abstract

Remote visual inspection is a common approach to understanding the health of key components and substructures within nuclear power plants, particularly in difficult to access and high dosage areas. Interpretation of inspection footage is a manually intensive procedure and challenges arise in localizing and dimensioning defects directly from a video feed, which may be subject to uncertainty from a range of sources such as lens distortion, nonuniform lighting, and lack of depth from a monocular camera system. A common approach to addressing these issues is to develop a scaling factor based on identifying a reference object of known dimensions in the image and using this to size regions of interest. Manual, accurate identification of these reference objects is onerous, time consuming, and prone to variation across different human experts, therefore, robust identification of suitable reference objects in an automated, reliable, and repeatable manner is of significant value. In this paper we evaluate two approaches for the automated detection of reference objects in the inspection of graphite cores in the United Kingdom’s fleet of advanced gas-cooled reactors (AGRs). The first method is a multistep approach using tools from mathematical morphology. The approach uses a genetic algorithm to “grow” suitable structuring elements, refine the order of operations, and remove operations proposed by the human designer that have a negative impact on performance. The second approach uses semantic segmentation, a technique which is normally applied to scene labeling in computer vision applications, applied to produce a binary mask, separating the reference object from the background. We show that this second method performs significantly better than the mathematical morphology approach when applied to the identification of brick interface keyways in AGR inspection images. Though improved in terms of accuracy, it is recognized that a greater initial effort is required to train the approach, and as it utilizes black-box neural network approaches, the greater transparency offered by the mathematical morphology approach is lost. While explicability of techniques is often a highly desirable characteristic of automated analysis techniques applied to health assessment within nuclear power plants, the results of the reference object detection can be made explicit to the end user, ensuring that the human analyst is retained within the decision-making process thus mitigating the need for transparency.

摘要

远程目视检查是了解核电厂关键部件和子结构健康状况的常用方法，尤其是在难以接近和高剂量区域。检查镜头的解释是一个人工密集的过程，在直接从视频输入中定位和确定缺陷尺寸时会出现挑战，这可能会受到一系列来源的不确定性的影响，例如镜头失真、不均匀的照明和单目相机缺乏深度系统。解决这些问题的常用方法是基于识别图像中已知尺寸的参考对象并使用它来确定感兴趣区域的大小来开发缩放因子。手动、准确地识别这些参考对象既费时又费力，并且容易因不同的人类专家而异，因此，以自动化、可靠和可重复的方式可靠地识别合适的参考对象具有重要价值。在本文中，我们评估了在英国先进气冷堆 (AGR) 中检测石墨堆芯时参考物体自动检测的两种方法。第一种方法是使用数学形态学工具的多步骤方法。该方法使用遗传算法来“增长”合适的结构元素，细化操作顺序，并删除人类设计者提出的对性能有负面影响的操作。第二种方法使用语义分割，这是一种通常应用于计算机视觉应用中场景标记的技术，用于生成二进制掩码，将参考对象与背景分开。我们表明，当应用于识别 AGR 检测图像中的砖界面键槽时，第二种方法的性能明显优于数学形态学方法。尽管在准确性方面有所提高，但人们认识到需要更大的初始努力来训练该方法，并且由于它利用了黑盒神经网络方法，数学形态学方法提供的更大的透明度会丢失。虽然技术的可解释性通常是应用于核电厂健康评估的自动化分析技术的一个非常理想的特征，但参考对象检测的结果可以向最终用户明确显示，确保人类分析员参与决策——从而减少对透明度的需要。