Guillaume Ricciardi

This study introduces AIA-Net, standing for Autoencoder Improved Attention Network, a novel neural-network (NN) architecture based on an attention mechanism for solving the non-linear inverse problem of Electrical Impedance Tomography (EIT). It aims at producing reliable reconstructed images of the conductivity distribution within a probed medium. To overcome the generalization problem inherent to deep learning approaches, AIA-Net is further improved using a latent space generative NN based on a Conditional Latent Diffusion Model. We present an end-to-end workflow from measurements to image reconstructions using AIA-Net. The EIT voltage excitations protocol considered relies on the ONe Excitation for Simultaneous High-speed Operation Tomography method. Static experimental tests are conducted, with plastic rods of different shapes and combinations emulating anomalous objects within a reference medium. Using a total of 16 320 data, comprising both voltage measurements and ground-truth images, some reference NNs for reconstructing conductivity images are deployed and compared, qualitatively and quantitatively, with the proposed AIA-Net architecture. The results presented highlight the excellent overall performance of AIA-Net compared to existing architectures, along with a further improvement provided by a generative model to overcome the generalization problem.

Electrical Impedance Tomography (EIT) offers spatially distributed, high-rate measurements of fluid properties in industrial pipelines, making it a promising sensing modality for early fault detection. However, anomaly detection in EIT signals faces specific challenges: the data are inherently unlabeled, the multivariate measurements exhibit complex cross-correlations, and baseline drift causes false alarms with threshold-based approaches. This article introduces Variational Unsupervised Self-Adversarial Detection (VUSAD), which combines variational autoencoder regularization with dual-decoder adversarial training to address these challenges. The variational component prevents overfitting to training samples, while the adversarial mechanism amplifies reconstruction errors for out-of-distribution inputs. Validation on a standard time-series benchmark demonstrates strong performance on industrial and environmental monitoring datasets, while revealing limitations on data with multimodal normal behavior. Application to EIT sensors monitoring fluid pipelines shows that the method can detect diverse fault types including cavitation, dry running, and foreign body passage with inference latency suitable for real-time deployment.

This study is concerned with the classification of two-phase flow regimes. The data considered are provided by electrical measurements at the boundary of the probed medium. In this context, our aim is to measure and characterize flow regimes, and their transitions, in real time. To this end, we investigate a neural-network approach based on spectrogram representations of time-dependent measurements in order to improve the accuracy compared to time series analyses. In the present study, dynamical tests in an experimental two-phase flow loop are considered, making use of a sensor developed at the Atomic Energy Commission coupled with a simultaneous voltage excitation method. A total of 80 configurations are studied, with different air/water mass flow rates leading to five different flow regime classes, namely slug, plug, wavy, stratified and annular. For each configuration, the available electrical measurements are processed into a normalized impedance matrix, whose important features are extracted using singular value decomposition. The time evolution of the maximum singular value is then passed to a neural network, which is trained from labeled data while making use of the stratified k -fold cross-validation method. Two different network architectures are considered: the first one processes raw time-series, leading to a regime classification accuracy of 89.4%. The second one uses spectrogram images, resulting in a more accurate classification reaching 97.4%. A comparison with other methods for identifying two-phase flows is also proposed, which supports the main contribution of this study in providing evidence of the advantages of using spectrograms of electrical impedance tomography measurements.

The design of advanced experimental nuclear reactors consists in integrating safety and operational requirements as well as reaching targets in terms of thermal power and neutron spectrum. In order to meet theses constraints, slender structures with little supports and crossing the entire reactor vessel are implemented in the reactor and are subjected to an axial flow that generates flow-induced vibration (FIV). From the industrial point of view, the mastering of the occurrence of FIV and its associated wear in case of contacts is requested. The stationary fluid forces applying on slender tubular structure in response to its motion are of primary interest since they can significantly affect the vibration amplitudes and even the stability of the system, especially in case of confined axial flow with high velocities (typically higher than 10 m/s). In this study it proposed to compare two approaches to estimate the vibration induced by turbulent excitation of an industrial device encountered in a research nuclear reactor. The control-rod guide-tube mock-up of the Jules Horowitz Reactor, previously tested in an hydraulic channel at the Technical Center from Le Creusot in France, is retained for this benchmark. Two models are proposed, one based on derivation of leakage flow theory and the other one was based on potential flow theory with adjusted coefficients given by CFD simulations. The flow induced vibration amplitude is consistent with the experimental data. Also, the calculation and experiment provide similar trends when the boundary conditions are changed.