A thermal monitoring system combining drones and deep neural networks can detect anomalies in nuclear power plants in real time, without exposing personnel to inaccessible or high-radiation areas. This was demonstrated by a research team from the Department of Nuclear Engineering at UNIST (Ulsan National Institute of Science and Technology) in South Korea, in a study published in the Journal of Nuclear Science and Technology in July 2025.
The system’s core lies in the integration of drone-mounted infrared imaging and deep learning techniques for object detection and segmentation. The drone flies over plant components, simultaneously capturing thermal and visual imagery, while an AI model analyzes the data to identify the location and severity of potential faults. Detection accuracy exceeds that of conventional visual inspection methods, with the added benefit of delivering real-time spatial information across components spread over large and physically complex areas.
The problem this approach addresses is a practical one. Nuclear power plants contain zones where direct human access is restricted or entirely off-limits due to radiation exposure. Conventional inspection methods — fixed sensors, scheduled manual inspections — provide neither continuous coverage nor the ability to pinpoint an anomaly with geometric precision. The UNIST system works on a regional basis: rather than measuring a single point, it maps an entire area, providing accurate indications of where a problem is located and how extensive it is.
The researchers trained the model on thermal images acquired under various operating conditions and simulated fault scenarios, using a small-scale integral hydraulic test facility. Convolutional neural networks learned to distinguish normal thermal variations from meaningful anomalies, even in cases where the temperature difference is minimal and barely perceptible to the human eye. Significant operational anomalies producing thermal variations of just a few degrees are correctly identified — a result that traditional manual analysis techniques cannot reliably match.
The operational implications are straightforward. A system of this kind can reduce plant downtime by enabling intervention before a minor fault becomes critical. It can also decrease the number of manual inspections in hazardous zones, directly lowering the cumulative radiation dose absorbed by workers over the course of their careers — no minor consideration, given that managing staff dose levels is a permanent operational constraint in nuclear facilities.
The next step is validation on full-scale real plants. Tests documented so far have been conducted at a small-scale experimental facility, and scaling up to a commercial reactor introduces additional variables — vibrations, electromagnetic interference, more complex geometries — that the system will need to prove it can handle. Should the technology clear that hurdle, drones and artificial intelligence could become a standard component of predictive maintenance protocols at next-generation plants, including the small modular reactors currently under development around the world.
