Self-Learning Acoustic Response Framework for Unsupervised Structural-State Identification in Metal Additively Manufactured Specimens

This study evaluates label-free clustering of active acoustic responses acquired under prescribed compressive-load conditions from fully debound and sintered 17-4 PH stainless-steel specimens fabricated by Metal Fused Filament Fabrication. One physical specimen was tested for each of three geometries: dog bone, I-beam, and obround plate. Surface Response to Excitation measurements were acquired using surface-mounted and post-fabrication cavity-mounted piezoelectric transducers. Six force conditions were applied sequentially from 0 to 500 N in 100 N increments, with 50 acoustic measurements collected while each load remained applied. These conditions represented prescribed force levels rather than manufacturing defects or verified damage states. Engineered three-channel spectrograms were evaluated using convolutional-autoencoder latent features followed by principal component analysis (PCA) and a protocol-matched direct-PCA baseline. Clustering was performed separately for 14 geometry–sensor–location combinations using automatically selected cluster counts and a specified count of six. Withheld load labels were introduced only after clustering to calculate the adjusted Rand index, normalized mutual information, and post hoc mapped clustering accuracy. Direct PCA achieved configuration-averaged mapped accuracies of 89.74% with automatic cluster-count selection and 98.52% with six specified clusters, compared with 85.23% and 91.42%, respectively, for the tested autoencoder pipeline. Direct PCA therefore provided stronger aggregate agreement with the prescribed force labels under the evaluated settings, while automatic recovery of six clusters remained configuration-dependent. These findings demonstrate within-specimen separation of the recorded responses. However, repeated acquisitions from one specimen per geometry do not establish specimen-to-specimen reproducibility, and sequential loading cannot fully exclude acquisition-order or loading-history effects. Cross-geometry generalization and practical deployment performance remain unverified.