Crack identification for marine engineering equipment based on improved SSD and YOLOv5. (15th January 2023)

• Record Type: Journal Article
• Authors: Jia, Ziguang; Su, Xin; Ma, Guangda; Dai, Tongtong; Sun, Jiabin
• Is Part Of: Ocean engineering. Volume 268(2023)
• Publication Date: 2023-01-15
• Subjects: FPSO -- Deep learning -- SSD -- YOLOv5 -- Attention mechanism; Ocean engineering -- Periodicals; Ocean engineering; Periodicals; 620.4162
• DOI: 10.1016/j.oceaneng.2022.113534
• Languages: English
• ISSNs: 0029-8018

Abstract

Abstract: The reliability requirements of marine engineering structures are becoming increasingly significant with the development of offshore oil and gas resources. This paper focuses on crack identification for floating production storage and offloading (FPSO) and makes the following research contributions. First, a dataset of 1360 low resolution crack images from FPSO is established, which may not be the best for crack identification. Then, two improved models based on deep learning algorithms are proposed, which are verified through experiments. The first model introduces an attention mechanism and deconvolution module that attain a high F1 score and a mAP index of 0.855 and 84.75%, respectively, which are 0.03 and 2.42% higher than the original model. In the second model, by introducing a dense connection block and adding another prediction head, the F1 score and mAP index reached 0.903 and 89.48%, respectively, which are 0.035 and 2.97% higher than the indices of the original model. These algorithms can be specifically used of for crack identification in ocean engineering environments to provide some new insights. Highlights: Made a dataset for FPSO, which is a good supplement to the dataset of Marine engineering and industrial defects. Two improved crack identification models for low resolution images under special conditions are propose. The F1 and mAP indexes of the first model are improved by 0.03 and 2.42% compared with the original model, respectively. The F1 and mAP indexes of the second model are improved by 0.035 and 2.97% compared with the original model, respectively.