Predictive nuclear power plant outage control through computer vision and data-driven simulation

Field operation and preparation (FO & P) processes in the outages of nuclear power plants (NPPs) involve tedious team coordination processes. This study proposed a predictive NPP outage control method through computer vision and data-driven simulation. The proposed approach aims at automatically detecting abnormal human/team behaviors and predicting delays during outages. Abnormal human/team behaviors, such as prolonged task completion and long waiting time, could induce delays. Timely capturing these field anomalies and precisely predicting delays is critical for guiding schedule updates during outages. Current outage control relies heavily on manual observations and experience-based field adjustments, which require extensive management efforts. Real-time field videos that capture abnormal human/team behaviors could provide information for supporting the prognosis of abnormal FO & P processes. However, manual video analysis could hardly provide timely information for diagnosing delays. Previous studies show the potentials of using real-time videos for capturing field anomalies. These studies fell short in examining automatic video analysis in compact work environments with significant occlusions. Besides, limited studies revealed how the captured field anomalies trigger delays during outages. Computer vision techniques have the potential for automating field video analysis and detections of prolonged task completions and long waiting times. This paper aims at automating the integrated use of 1) real-time computer vision and spatial analysis algorithms, and 2) data-driven simulations of FO & P processes for supporting predictive outage control. The authors first use the video-based human tracking algorithm to detect human/team behaviors from field videos. Then, the authors formalized detailed human-task-workspace interactions for establishing a simulation model of FO & P processes during outages. The simulation model takes the field anomalies captured from videos as inputs to adjust model parameters for achieving reliable predictions of workflow delays. Major observations show that 1) task delays often occur at the initial stage of the workflow, and 2) waiting line accumulates due to excessive resource sharing during handoffs at the middle stage of the workflow. The simulation results show that tasks on the critical-path are more sensitive to these anomalies and cause up to 5.53% delays against the as-planned schedule.