Damage detection for tethers of submerged floating tunnels based on convolutional neural networks. (15th April 2022)

• Record Type: Journal Article
• Authors: Min, Seongi; Jeong, Kiwon; Noh, Yunhak; Won, Deokhee; Kim, Seungjun
• Is Part Of: Ocean engineering. Volume 250(2022)
• Publication Date: 2022-04-15
• Subjects: Submerged floating tunnel -- Damage detection -- Convolutional neural networks -- Mooring -- Tether; Ocean engineering -- Periodicals; Ocean engineering; Periodicals; 620.4162
• DOI: 10.1016/j.oceaneng.2022.111048
• Languages: English
• ISSNs: 0029-8018

Abstract

Abstract: A submerged floating tunnel, consisting of a tunnel and tethers, is effective as a sea-crossing transportation infrastructure element in deep-water environments. Instead of fixed columns, the tunnel is positioned by the tethers. This means that a significant structural performance degradation in the mooring can directly induce a change in the structural state; moreover, the failure of the tethers will eventually lead to structural instability. Therefore, structural health monitoring is essential for the tethers, as well as for the main tunnel segments. Unfortunately, there are limitations to the applicable sensors for measuring the structural responses required to evaluate the structural state and for estimating the structural damage to the tethers, owing to the environmental characteristics. Therefore, it is necessary to develop and apply an effective damage detection method to secure structural safety. Accordingly, in this study, an advanced damage detection method is proposed for the tethers of submerged floating tunnels based on the convolutional neural network (CNN). The damage detection estimation model is based on a convolutional neural network framework consisting of input, output, and hidden layers for training, validation, testing, and application. The model is trained using structural response data obtained by a hydrodynamics-based time-domain analysis considering various waves and tether damage cases. For successful training, the time-domain structural response data are converted to discretized image data. The accuracy of the proposed CNN-based damage detection models for the various damage rates and noise levels was evaluated. The accuracy of the CNN–S-16-Model which uses 16 sensors with 0–5% level noise signals ranges from 98.4 to 100.0% for 50% damage, from 97.6 to 100.0% for 30% damage, and from 92.0 to 100.0% for 15% damage to the tethers. The noise level significantly affected the damage detection accuracy for the relatively low damage rate cases. Therefore, rational noise filtering is required to enhance the accuracy for minor damage cases. Highlights: CNN-based damage detection method for SFT tethers is suggested. For training the damage detection model, the acceleration data measured inside the SFT under various wave conditions is used. The detection accuracy is affected by the damage rate and assumed noise level. The lowest damage detection accuracy is 91.54% for the 15% damage cases with a 5% noise level.