TY  - EJOU
AU  - Zhang, Boyang 
AU  - Gao, Kang 
AU  - Gu, Jianan 
AU  - Ai, Li 
AU  - Geng, Yuang 

TI  - Improving Stage-Transition Detection in Long-Term Guided-Wave Monitoring through Autoencoder Regularization
T2  - Structural Durability \& Health Monitoring

PY  - 
VL  - 
IS  - 
SN  - 1930-2991

AB  - Long-term guided-wave structural health monitoring accumulates large volumes of sensing data, but dense damage labels are rarely available in field deployment. Unsupervised reconstruction-based detection methods, which identify damage through reconstruction discrepancy, are therefore appealing for their ability to learn directly from unlabeled data. However, environmental and operational variations and mixed adjacent-stage samples in transition windows both degrade unsupervised detection performance. Regularization offers a principled way to improve the robustness of unsupervised detection against such confounding variations, but its effect on fine-grained, transition-level detection has not been systematically studied. This paper presents a controlled empirical study to evaluate various regularization methods for unsupervised transition-level damage detection under uncontrolled environmental conditions. Using the public Scientific Data benchmark, four regularization configurations, namely unregularized, dropout, weight decay, and sparsity, are compared under an identical architecture and training protocol. Detection performance is evaluated at adjacent stage transitions, which reflect the practical question of whether a new structural state differs from the immediately preceding one. Pearson correlation coefficient (PCC) trajectories and area under the receiver operating characteristic curve (AUROC) serve as the primary metrics, and a transition is counted as detected only when both AUROC <mml:math id="mml-ieqn-1"><mml:mo>≥</mml:mo></mml:math> 0.85 and average PCC gap between two states <mml:math id="mml-ieqn-2"><mml:mo>≥</mml:mo></mml:math>0.010 are satisfied. Under this criterion, the unregularized autoencoder detects 2 of 12 adjacent transitions, while dropout, weight decay, and sparsity detect 5, 6, and 7 transitions, respectively, with best mean AUROC values of 0.9071, 0.9249, and 0.9288. Analysis of representative difficult cases reveals two distinct failure regimes: weak structural contrast between adjacent stages and environment-dominated signal masking. These results demonstrate that regularization alone, without architectural modification, can materially increase transition-level detection coverage in long-term uncontrolled guided-wave monitoring.
KW  - Guided-wave structural health monitoring; unsupervised damage detection; autoencoder regularization; uncontrolled environment; adjacent stage-transition detection; reconstruction correlation

DO  - 10.32604/sdhm.2026.083281