Unsupervised novelty detection for EMAT-based inspection on embedded hardware

The interpretation of Electromagnetic Acoustic Transducer (EMAT) signals for assessing corrosion-induced wall loss in metallic plates are constrained by the scarcity of labeled defects in industrial operation, motivating one-class formulations. To address this, the 20,000-sample A-scan undergoes reduction to a 23-dimensional RMS envelope using band-pass filtering, dead-zone suppression, and sequential windowing. Next, an autoencoder, an Isolation Forest, and a centroid-distance KMeans are trained on normal specimens, and thresholded under a common three-sigma rule applied to their respective training scores. Finally, embedded inference on a Raspberry Pi 4 yields 99% specificity and a near-unity AUC-ROC for both throughwall holes and three graded levels of thickness loss. Five-fold cross-validation confirms low dispersion (standard deviation of 0.0% for for sensitivity and specificity), supporting stability under limited training data.