Machine-Learning Evaluation of a Magnetostrictive Acoustic-Emission Sensor Developed for Nuclear Structural Health Monitoring

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Acoustic emission (AE) sensing is widely used for structural health monitoring, but conventional piezoelectric sensors can be constrained in high-temperature and radiation environments. This study evaluates whether an Idaho National Laboratory-developed magnetostrictive AE sensor retains sufficient information for automated impact-source classification under a controlled room-temperature configuration. A total of 334 synchronized acquisitions from six impact classes were recorded at 20 MHz using three fixed, non-coincident sensor channels. Because sensor type was not independently varied from position, mounting, coupling, bandwidth, or propagation path, comparisons represent complete sensor-channel responses rather than isolated transduction mechanisms. Each waveform was represented by 16 time-domain, spectral, and band-power features. Random forest classifiers were evaluated for individual channels, 48-feature fusion, and equal-vote decision fusion using 30 repeated stratified holdout runs and stratified 10-fold cross-validation. The magnetostrictive channel achieved 95.35 ± 2.45% mean holdout accuracy, compared with 92.37 ± 2.94% and 96.31 ± 1.98% for the two commercial channels. Feature-level fusion achieved 97.12 ± 2.19% and 97.91 ± 3.18% under holdout and 10-fold cross-validation, respectively. Because class-specific blocks were split at the event level, these accuracies are within-campaign and within-specimen, may be optimistic, and do not establish block-independent or operational generalization; validation using randomized or interleaved repeated blocks, multiple specimens, position-controlled comparisons with swapped, rotated, or co-located sensor placements, and harsh-environment testing remains necessary.