Scholarlink

<p>The development of integrated artificial intelligence algorithms for creating effective and efficient structural health monitoring platforms has seen significant advancements and applications over recent decades. However, as with the introduction of new technologies to any domain, there are unique challenges for applying artificial intelligence models for damage detection, classification, localization, and quantification of structures in structural health monitoring. In particular, the logistical, economic, and safety restraints associated with the extraction of data from structures result in datasets that experience limited, unbalanced class samples, resulting in poorly trained and biased AI models that perform sub-optimally. Though extensive research has been conducted in the domain of data augmentation, whereby new samples containing similar statistical information as existing data are created, this area of study faces its limitations, including the quality and diversity of generated samples, and the effectiveness of previously developed techniques. In this study, a light gradient boosting model is used to rank the application of various data augmentation techniques on existing time and frequency-domain features extracted from 1D time-series structural data. Through the prediction of a novel predictive indicator, the area in the radar plot, the light gradient boosting model can determine the performance of a data augmentation technique corresponding to a selected classifier used for the detection and classification of damage characteristics. The proposed model was validated by extensively implementing 5 different data augmentation algorithms and 5 different machine learning and deep learning techniques applied to 3 different frame-based structures experiencing various structural damages. The aggregated results demonstrated considerable accuracy at predicting the correct ARP value and ranking of DA techniques on a given dataset, provided that the dataset’s information was represented in both the training and testing subsets (MSE = 0.002, MAE = 0.024, RMSE = 0.026, SMAPE = 3.31%). This model pipeline and the development of future DA optimization algorithms could further address limitations in the inspection-based SHM domain applications by autonomously expanding limited, unbalanced vibrational data extracted from structures.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Light gradient boosting machine for data augmentation optimization and model selection for one-dimensional multiclass damage detection

  • 期刊论文

  • 发表日期：

    2026年7月21日

查看全文
  • Kyle Dunphy,
  • Zachary Baird,
  • Mohamed Barbosh,
  • Ayan Sadhu

摘要

The development of integrated artificial intelligence algorithms for creating effective and efficient structural health monitoring platforms has seen significant advancements and applications over recent decades. However, as with the introduction of new technologies to any domain, there are unique challenges for applying artificial intelligence models for damage detection, classification, localization, and quantification of structures in structural health monitoring. In particular, the logistical, economic, and safety restraints associated with the extraction of data from structures result in datasets that experience limited, unbalanced class samples, resulting in poorly trained and biased AI models that perform sub-optimally. Though extensive research has been conducted in the domain of data augmentation, whereby new samples containing similar statistical information as existing data are created, this area of study faces its limitations, including the quality and diversity of generated samples, and the effectiveness of previously developed techniques. In this study, a light gradient boosting model is used to rank the application of various data augmentation techniques on existing time and frequency-domain features extracted from 1D time-series structural data. Through the prediction of a novel predictive indicator, the area in the radar plot, the light gradient boosting model can determine the performance of a data augmentation technique corresponding to a selected classifier used for the detection and classification of damage characteristics. The proposed model was validated by extensively implementing 5 different data augmentation algorithms and 5 different machine learning and deep learning techniques applied to 3 different frame-based structures experiencing various structural damages. The aggregated results demonstrated considerable accuracy at predicting the correct ARP value and ranking of DA techniques on a given dataset, provided that the dataset’s information was represented in both the training and testing subsets (MSE = 0.002, MAE = 0.024, RMSE = 0.026, SMAPE = 3.31%). This model pipeline and the development of future DA optimization algorithms could further address limitations in the inspection-based SHM domain applications by autonomously expanding limited, unbalanced vibrational data extracted from structures.