Enhanced video-level anomaly feature detection for nuclear power plant component inspections using the latency mechanism

Inspections are regularly carried out at various nuclear power plant components to provide latest understanding of the facility condition. One common approach is via remote visual inspection which requires the engineers to watch a large volume of video footage and identify the anomalies in the video. This process is intensively manual-based as the video segments containing anomalies of interest is only a small part of the original video. For this reason, an automated anomaly detection tool is expected to ease the amount of human effort involved in the inspection process. Deep learning is a useful tool to autonomously detect anomalies in the inspection video, given that a well-prepared training dataset of anomaly types is available. However, the detection system may detect false-positives which can be difficult to remove without human intervention. To solve this problem, we introduce a new video-level detection workflow incorporating the latency mechanism to effectively reduce the false-positive detections in the inspection videos. In this workflow, each region in a frame is scanned using a convolutional neural network (CNN) trained for a specific anomaly type. The initial scanning results are then refined using the latency mechanism which flags a region as “anomaly” when the anomaly is detected in the current frame and in a series of previous consecutive frames. A case study of crack-like feature detection in superheaters is presented to demonstrate the efficacy of the proposed workflow. The results show that the false-positive detections seen in the initial scanning can be effectively reduced using the proposed latency mechanism. It is suggested that this workflow can be directly transferable to various anomaly detection tasks of nuclear plant facility inspection.