Session: 06-02: Machine Learning and Statistical Methods in NDE II

Paper Number: 144657

144657 - Improving Steam Generators Nuclear Power Plant Inspection Through Ai Using Eddy Current Ndt Data 

Abstract: 

Authors: Othmane ACHOUHAM¹, Charfeddine MECHRI¹, Rachid EL GUERJOUMA¹, Christian NOEL², Jimmy PONTON²

¹LAUM - Le Mans University - CNRS – UMR 6613 – Le Mans - FRANCE

²OMEXOM NDT Engineering & Services – Bourg de Péage - FRANCE

Steam generators are critical components of nuclear power plants, playing a vital role in facilitating heat transfer between the primary and secondary circuits. However, the presence of defects such as corrosion, cracking, or thinning of tube walls can compromise their mechanical integrity and thermal efficiency, posing a significant risk to the efficiency, safety, and reliability of the facilities. Consequently, non-destructive testing (NDT) of steam generators is of paramount importance, requiring precise detection of potential defects in the tubes. Among the most commonly used methods, eddy current, and for more specific cases ultrasonic testing stand out for their effectiveness in inspecting these tubes.

However, beyond the intrinsic efficiency of these control methods, the challenge lies in the need to minimize downtime and associated inspection costs, especially when hundreds of kilometers of tubes require regular monitoring to ensure an optimal level of safety. Faced with the rapidly increasing number and frequency of inspections, along with the growing amount of collected data, the objective of our study is to facilitate and expedite the work of inspectors in the field by automatically identifying defect-free areas. This allows inspectors to focus on analyzing potentially problematic zones. The use of artificial intelligence, capable of rapidly processing vast sets of complex and heterogeneous data, could achieve this goal by automating the analysis process, ultimately accelerating inspection frequencies and reducing the manual workload of analysts in the field.

In our study, we focus on utilizing data from eddy current inspections using multifrequency probes. These probes provide multiple frequency responses for each measurement point, organized into detection channels. Since defects are rare in steam generators, defect signals represent less than 0.8% of the total database, resulting in a highly unbalanced distribution between defect and healthy signals. This imbalance renders supervised machine learning approaches less effective, prompting us to adopt an unsupervised approach involving anomaly detection algorithms. Leveraging the prevalence of healthy signals in the database, the algorithm can learn to detect the characteristics of healthy signals and identify those deviating from the norm. To fuel our anomaly detection models, we conducted indicator extraction, identifying 38 relevant indicators for each frequency channel based on literature. A feature selection step was then implemented to expedite and optimize the detection procedure.

During the inspection of tubes in steam generators, no sub-detection is permissible, meaning the detection of a defect must not be missed. Thus, the evaluation of our algorithm is based on zero sub-detection (false negatives) and a minimal rate of over-detection (false positives). This approach yielded highly satisfactory results. Utilizing the labeling provided by the control operator, we evaluated the performance of our algorithms and optimized their ability to automatically identify areas for inspection, thereby facilitating and accelerating the work of field analysts. The deployment of artificial intelligence for these inspections will enhance analytical capabilities, improving the reliability of result interpretations.

This project is part of the AUTEND project funded by "FRANCE RELANCE" and "FRANCE 2030" programs.

Presenting Author: Othmane Achouham LAUM - Le Mans Université - CNRS

Presenting Author Biography: I am Othmane ACHOUHAM, a second-year PhD student at the Acoustics Laboratory of the University of Le Mans. My doctoral research focuses on improving the inspection process of steam generators in nuclear power plants by analyzing eddy current non-destructive testing (NDT) data using artificial intelligence techniques.

Authors:

Othmane Achouham LAUM - Le Mans Université - CNRS
Rachid El Guerjouma LAUM - Le Mans Université - CNRS
Charfeddine Mechri LAUM - Le Mans Université - CNRS
Christian Noel OMEXOM NDT Engineering & Services
Jimmy Ponton OMEXOM NDT Engineering & Services