Convolutional neural network-based safety evaluation method for structures with dynamic responses

• Published in: Expert systems with applications 2020-11, Vol.158, p.113634, Article 113634
• Main Authors: Seon Park, Hyo, Hwan An, Jung, Jun Park, Young, Kwan Oh, Byung
• Format: Article
• Language: English
• Subjects: Artificial neural networks; Convolutional neural network; Dynamic structural response; Malfunctions; Neural networks; Performance prediction; Reinforced concrete; Safety evaluation; Sensors; Shake table tests; Strain measurement; Strain monitoring; Structural health monitoring; Structural members; Structural safety
• ISSN: 0957-4174; 1873-6793
• DOI: 10.1016/j.eswa.2020.113634

•A convolutional neural network-based strain estimation method is presented.•The networks identify the relationship between dynamic response and strain.•Constructed networks evaluate structural safety by estimating strain response.•Effectiveness of the method is validated by both numerical and experimental studies. The strain sensors that are used to evaluate structural members have a limited lifespan and thus have shown limitations to perform long-term structural health monitoring (SHM). This study presents a convolutional neural network (CNN)-based strain prediction technique that allows for structural safety evaluations in case of absence or defect of strain sensors. In the proposed method, CNNs were used to establish a relationship between the dynamic structural response and the strain response measured in the structure. A number of dynamic structural responses and the structural member’s strain response that are measured before the strain sensor malfunctions are used as input data and output data, respectively, to train the CNNs. The trained CNNs can estimate the strain and evaluate the structural safety even when the later strain measurement response cannot be used. Dynamic acceleration and displacement responses are used as input data in the two CNNs presented in this study, called CNN_A and CNN_D respectively. A numerical study of a beam-like structure and an experimental study which includes shaking table tests on a reinforced concrete frame specimen were conducted to confirm the validity of the strain predictions by the proposed method with CNN_A and CNN_D. The strain prediction performance of the proposed CNNs is compared in these applications. This study also examines the proposed technique’s strain prediction performance according to the amount of data used to train the CNNs. In addition, this study discusses influences of variations in the number of locations for measuring the dynamic structural responses that are used as the CNNs’ input data on the strain prediction performance.