Abstract
The environment inside Fukushima Daiichi is highly unstructured with unknown elements. There is not sufficient CAD data available of structure or equipment inside the reactor, and knowledge of contents may be missing or unreliable due to limited data following the disaster and subsequent activities. These factors limit options for object classification.
This paper will present the research work completed by LongOps programme to develop a dataset for an applied machine learning model to perform object detection and segmentation on the structures and contents inside the Primary Containment Vessel (PCV) (Tokyo Electric Power Company Holdings Inc. Official Website, 2023). To build the dataset, annotation, process to define boundary of an object and label it with corresponding class on an image or video, has been performed using past PCV investigation videos. This task was challenging by the presence of noise and haze caused by the water and propeller motion of the remotely operated vehicle carrying the camera(s). General classes were used to cater for the wide range of damaged equipment and structures. In total, 13 classes of components and structures including corrosion and fuel debris have been identified. To cater to the low number of annotations of some classes (also known as imbalanced dataset), new images have been created by adding noise, rotation, zoom, blurriness, and flipping to original images from PCV videos to improve the data imbalance between classes.
This study work also demonstrates the capability of detecting and segmenting components using a machine learning model. For this purpose, a Mask Region-based Convolutional Neural Network model has been used for the dataset that has been split into three categories: train, validation, and a test dataset to test the performance of the trained model. For safe operation of the robot(s), corrosion can be detected so that robot can be moved safely without further damaging the structures. In addition, fuel debris can also be detected in current trained model to help the operator in locating and removing them in a safe and efficient way.
