First cycle Last cycle Evolution of the cooling down time First cycle Last cycle LBP histogram LBP histogram Failure detection Failure classification Remaining useful life (RUL) In the context of Reliability, Availability, Maintainability, and Inspectability (RAMI) analysis for DEMO, achieving an availability of 30-60% and minimizing unscheduled shutdowns are essential prerequisites. Predictive maintenance (PdM) can play a crucial role in ensuring regular, rapid, and reliable maintenance of the plant. Critical components, such as the divertor and the first wall, face extreme thermal loads, intense thermal shocks, and bombardment by plasma ions, neutral particles, and energetic neutrons. The options for real-time monitoring of Plasma-Facing Components (PFCs) in such conditions are limited. Moreover, computational models for reactor-scale surfaces can require significant computational resources. To overcome these challenges, we employ data-driven PdM to analyze infrared data and domain expert annotations, identifying patterns and anomalies that may indicate damage to the PFC. In particular, this study focuses on the application of modern statistical methods, including machine learning and deep learning, on infrared data from steady-state heat load experiments. Predictive maintenance in fusion devices with an application to condition monitoring of plasma-facing components Efficiently Optimal operations Safely No catastrophic failures Reliably No downtime Fusion power plants will need to run: Detection, diagnosis, and prediction of failures in machine subsystems in order to prevent unexpected downtime What is predictive maintenance? Why? Predictive maintenance (PdM) is an approach that can contribute to meeting this requirement through periodic or continuous monitoring of equipment condition. The goal is to predict when maintenance will be needed for the equipment and, ultimately, provide an estimate of the remaining useful lifetime of devices and components. Real-time data To monitor the condition of the equipment Data-driven PdM Past historical data To represent the health status of the equipment L. Caputo, D. Dorow-Gerspach, M. Wirtz G. Verdoolaege 1 2 2 1 Department of Applied Physics, Ghent University, B-9000 Ghent, Belgium Forschungszentrum Jülich, Institut für Energieund Klimaforschung - Plasmaphysik, 52425 Jülich, Germany 1 2 Experimental setting: tiles exposed to an electron beam Evidence of tile failure from infrared data Evolution of the pixel intensities by k-means algorithm Anomaly detection using autoencoder neural network architecture An autoencoder neural network performs anomaly detection in the infrared data. The autoencoder, consisting of an encoder and a decoder, is trained on data representing normal, undamaged tile conditions. The encoder compresses the input data into a lowerdimensional representation, while the decoder reconstructs the original data from this representation. Anomalies are identified by significant reconstruction errors, indicating deviations from normal thermal behavior, which may be linked to damage in the tiles. Beryllium tiles are cyclically exposed to an electron beam to simulate a steady-state heat load, with power gradually increased until failure occurs. Thermal images are captured during each load cycle using an infrared camera. Excessive cooling down time Anomalies in image texture properties K-Means clustering identifies the hottest and coldest pixel clusters. The distance between these cluster centroids evolves over cycles, highlighting significant temperature changes that indicate potential overheating and damage. Two types of failures Surface temperature > 700 C Delamination o First cycle Last cycle Local Binary Pattern (LBP) is a type of visual descriptor used for classification in computer vision. Anomalies in the LBP histogram can indicate potential damage to the tile. The cooling down time of the tile within each cycle, as estimated from the thermal images, provides clear evidence of tile damage. Frame index Frame index Reconstruction error (MSE) Image index Contact Infusion, Department of Applied Physics, Ghent University 📧
[email protected]