{"id":1291,"date":"2025-07-01T15:58:08","date_gmt":"2025-07-01T15:58:08","guid":{"rendered":"https:\/\/smartdata.ece.ufl.edu\/?page_id=1291"},"modified":"2026-04-05T04:43:28","modified_gmt":"2026-04-05T04:43:28","slug":"publications","status":"publish","type":"page","link":"https:\/\/smartdata.ece.ufl.edu\/index.php\/publications\/","title":{"rendered":"Publications"},"content":{"rendered":"\n

SmartDATA Lab Publications<\/h2>\n\n\n
<\/a>
<\/div><\/div><\/form>
203 entries<\/span> «<\/a> ‹<\/a> 1 of 11 ›<\/a> »<\/a> <\/div><\/div>

2026<\/h3>
203.<\/div>
<\/div>

Kang Yang, Zekun Yang, Zhihui Tian, Harsha Tetali, Joel B Harley<\/p>

Weight decay optimized unsupervised autoencoder based anomaly \r\n detection in uncontrolled dynamic structural health monitoring<\/a> Book Section<\/span> <\/p>

In: <\/span>Lecture Notes in Computer Science, <\/span>pp. 31\u201339, <\/span>Springer Nature Switzerland, <\/span>Cham, <\/span>2026<\/span>.<\/p>

Abstract<\/a><\/span> | Links<\/a><\/span> | BibTeX<\/a><\/span><\/p>

@incollection{Yang2026-es,
\r\ntitle = {Weight decay optimized unsupervised autoencoder based anomaly 
\r\n detection in uncontrolled dynamic structural health monitoring},
\r\nauthor = {Kang Yang and Zekun Yang and Zhihui Tian and Harsha Tetali and Joel B Harley},
\r\nurl = {http:\/\/dx.doi.org\/10.1007\/978-3-031-94895-4_3},
\r\ndoi = {10.1007\/978-3-031-94895-4_3},
\r\nyear  = {2026},
\r\ndate = {2026-01-01},
\r\nbooktitle = {Lecture Notes in Computer Science},
\r\npages = {31\u201339},
\r\npublisher = {Springer Nature Switzerland},
\r\naddress = {Cham},
\r\nseries = {Lecture notes in computer science},
\r\nabstract = {Current autoencoder-based unsupervised anomaly detection in 
\r\n Dynamic Data Driven Applications Systems (DDDAS) typically 
\r\n depends on a comprehensive collection of historical normal 
\r\n signals as training data. However, such unsupervised models 
\r\n struggle to perform effectively under dynamic conditions when 
\r\n there is a significant divergence between the environmental 
\r\n conditions of the evaluation data and those of the training data. 
\r\n To address this, our previous study introduces an unsupervised 
\r\n anomaly detection method trained solely on the current evaluation 
\r\n data, eliminating the dependency on historical data and thus more 
\r\n applicable in DDDAS. This method utilizes the inherent bias 
\r\n learning property of neural networks, which typically prioritizes 
\r\n learning from larger classes (normal signals) at the expense of 
\r\n smaller ones. However, a crucial limitation of this new 
\r\n autoencoder-based damage detection method is its performance 
\r\n degradation when the evaluation data includes an increasing 
\r\n number of abnormal signals. To enhance anomaly detection under 
\r\n these conditions, an optimal weight decay regularization strategy 
\r\n is provided in this study to limit the autoencoder\u2019s ability to 
\r\n learn abnormal signals. The efficacy of the novel 
\r\n autoencoder-based anomaly detection method in DDDAS is validated 
\r\n using guided waves gathered from a structural health monitoring 
\r\n system under uncontrolled and dynamically varying environmental 
\r\n conditions.},
\r\nkeywords = {},
\r\npubstate = {published},
\r\ntppubtype = {incollection}
\r\n}
\r\n<\/pre><\/div>
Close<\/a><\/p><\/div>
Current autoencoder-based unsupervised anomaly detection in 
\r\n Dynamic Data Driven Applications Systems (DDDAS) typically 
\r\n depends on a comprehensive collection of historical normal 
\r\n signals as training data. However, such unsupervised models 
\r\n struggle to perform effectively under dynamic conditions when 
\r\n there is a significant divergence between the environmental 
\r\n conditions of the evaluation data and those of the training data. 
\r\n To address this, our previous study introduces an unsupervised 
\r\n anomaly detection method trained solely on the current evaluation 
\r\n data, eliminating the dependency on historical data and thus more 
\r\n applicable in DDDAS. This method utilizes the inherent bias 
\r\n learning property of neural networks, which typically prioritizes 
\r\n learning from larger classes (normal signals) at the expense of 
\r\n smaller ones. However, a crucial limitation of this new 
\r\n autoencoder-based damage detection method is its performance 
\r\n degradation when the evaluation data includes an increasing 
\r\n number of abnormal signals. To enhance anomaly detection under 
\r\n these conditions, an optimal weight decay regularization strategy 
\r\n is provided in this study to limit the autoencoder\u2019s ability to 
\r\n learn abnormal signals. The efficacy of the novel 
\r\n autoencoder-based anomaly detection method in DDDAS is validated 
\r\n using guided waves gathered from a structural health monitoring 
\r\n system under uncontrolled and dynamically varying environmental 
\r\n conditions.<\/div>
Close<\/a><\/p><\/div>