A novel unsupervised real‐time damage detection method for structural health monitoring using machine learning. Issue 10 (22nd July 2022)

• Record Type: Journal Article
• Authors: Shi, Sheng; Du, Dongsheng; Mercan, Oya; Kalkan, Erol; Wang, Shuguang
• Is Part Of: Structural control and health monitoring. Volume 29:Issue 10(2022)
• Page Start: n/a
• Page End: n/a
• Publication Date: 2022-07-22
• Subjects: convolutional neural networks -- deep support vector data description -- IASC–ASCE benchmark -- machine learning -- structural health monitoring -- unsupervised damage detection; Structural engineering -- Periodicals; Structural control (Engineering) -- Periodicals; Automatic data collection systems -- Periodicals; Detectors -- Periodicals; 624.17
• DOI: 10.1002/stc.3042
• Languages: English
• ISSNs: 1545-2255

Abstract

Summary: Real‐time structural damage detection is one of the main goals of establishing an effective structural health monitoring system. However, due to the lack of training data for possible damage patterns, supervised methods tend to be difficult for such applications. This article therefore proposes a novel unsupervised real‐time damage detection method using machine learning, which consists of a statistical modeling approach using neural networks and a decision‐making process using deep support vector domain description. To choose an optimal window length while extracting damage‐sensitive features, an iterative training strategy is proposed to remove redundant samples from an oversized window. The proposed method is then verified using a simulated dataset from the International Association for Structural Control–American Society of Civil Engineering benchmark and an experimental dataset from shake table tests. The results show that the mean alarm density can be used as an indicator of damage existence and damage levels for the single‐sensor approach. Higher performance of damage detection and lower performance of identifying damage levels are observed for the multi‐sensor approach when the rotational modes are amplified by asymmetric damage patterns. The results of mean false alarm density show that the presented method has a low probability of generating false alarms. The effectiveness of iterative pruning strategy is observed through the visualization of loss function and weights in the neural networks. Finally, the capability of real‐time execution of the proposed damage detection method is investigated and verified. As a result, trained with healthy data only, the proposed method is effective in detecting damage existence and damage levels.