期刊
JOURNAL OF VIBRATION AND CONTROL
卷 -, 期 -, 页码 -出版社
SAGE PUBLICATIONS LTD
DOI: 10.1177/10775463231209008
关键词
Hybrid machine learning; cold start; mode shape; principal component analysis; structural damage identification
This paper proposes an integration strategy of a modal-based method into a hybrid machine learning-based method for damage detection and localization of beam-like structures. By combining unsupervised and supervised methods, the cold-start and manual labeling issues are addressed, and high accuracy results are achieved.
Data-driven damage detection methods are widely researched and implemented due to the availability of advanced sensing and cloud technologies, where machine learning models are used to process the various data collected from the sensors for damage diagnosis. Supervised methods have shown to be accurate in identifying damages given that they are trained with sufficient historical damage samples. As damage events are rare, supervised methods often face the cold-start problem, that is, insufficient labelled damage samples to initiate model training. Unsupervised methods do not require training and have shown to be practical up to detecting damage presence, but prior knowledge or labelled historical damage samples are still required as references to further locate and quantify the damage. To solve the cold-start issue, a modal-based method, which relates the modal parameters with damages, is proposed to bridge the gap between the unsupervised and supervised methods. Therefore, this work proposes an integration strategy of a modal-based method into a hybrid machine learning-based method for damage detection and localisation of beam-like structures. A two-stage scheme with unsupervised, supervised, and modal-based methods is proposed. The first stage detects damage presence using PCA-FRF as the damage-sensitive feature in an unsupervised manner, and the second stage locates the damages using the first mode shape. Mode shape assessment-based damage localisation is performed at the second stage when no trained model is available due to zero or limited damage samples for training to solve the cold-start and manual labelling issue. Supervised model training starts once there are at least two classes of damage, and the trained model is only used for damage localisation when it surpasses the 90% validation accuracy benchmark. Results showed accuracies of 100%, 100%, and 95.85% for the first stage unsupervised detection, second stage mode shape assessment-based damage localisation and the supervised model-based damage localisation, respectively.
作者
我是这篇论文的作者
点击您的名字以认领此论文并将其添加到您的个人资料中。
推荐
暂无数据