A system reliability approach to real-time unsupervised structural health monitoring without prior information. (15th May 2022)

• Record Type: Journal Article
• Authors: Soleimani-Babakamali, Mohammad Hesam; Sepasdar, Reza; Nasrollahzadeh, Kourosh; Sarlo, Rodrigo
• Is Part Of: Mechanical systems and signal processing. Volume 171(2022)
• Publication Date: 2022-05-15
• Subjects: Unsupervised real-time SHM -- Generative adversarial networks -- Gaussian mixture models -- Anomaly detection -- System reliability -- Monte Carlo histogram sampling; Structural dynamics -- Periodicals; Vibration -- Periodicals; Constructions -- Dynamique -- Périodiques; Vibration -- Périodiques; Structural dynamics; Vibration; Periodicals; 621
• DOI: 10.1016/j.ymssp.2022.108913
• Languages: English
• ISSNs: 0888-3270

Abstract

Abstract: Unsupervised techniques have gained much attention in the last decade as the most practical real-time structural health monitoring (SHM) approach. However, there are still obstacles to robust real-time health monitoring among the proposed unsupervised methods in the literature. These barriers include loss of information from dimensionality reduction, case-dependency of feature extraction steps, lack of dynamic-class novelty detection approaches, and detection results' sensitivity to user-defined detection parameters. This study introduces an unsupervised real-time SHM method, addressing the above four obstacles simultaneously. Furthermore, the proposed technique requires no prior information from structures before going online as a step towards having a general SHM technique. No prior information is common for newly detected novelties, preventing the establishment of detection baselines. Without the baselines, dynamic-class novelty detection cannot take place. Hence, while solving the dynamic-class novelty detection hindrance with Generative Adversarial Networks (GAN), the framework can be adjusted to be prior-information-free. Online generations of data objects with GAN from the incoming data stream is the key feature that addresses the obstacles above. A mixture of low- and high-dimensional features are used to train multi-ensembles of GAN and one-class joint Gaussian distribution models (1-CG). A novelty detection system of limit-state functions based on GAN and 1-CG models' detection scores is constructed. The Resistance of the limit-state functions is tuned to user-defined detection parameters with the GAN-generated data objects through reliability analysis. The tuning makes the detection results robust to the user-defined detection parameters. The proposed novelty detection framework is applied to two standard SHM datasets to illustrate its generalizability: Yellow Frame and Z24 Bridge. All different damage categories are identified with low sensitivity to the initial choice of detection window length by employing the proposed dynamic and static baseline approaches with few or no false alarms. Graphical abstract: Highlights: Raw high-dimensional features are used for unsupervised structural health monitoring. Generative adversarial networks (GAN) are used to handle the high-dimensional features. Data generation with GAN mitigated the need for excessive prior information. Method performs dynamic-class novelty detection, distinguishing them from each other. Reliability analysis stabilizes performance per users' choice of hyperparameters. The framework is applied to a frame and a bridge dataset.