Premium
Waveform‐based fracture identification of steel beam ends using convolutional neural networks
Author(s) -
Wang Luyao,
Dang Ji,
Wang Xin,
Shrestha Ashish
Publication year - 2021
Publication title -
structural control and health monitoring
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.587
H-Index - 62
eISSN - 1545-2263
pISSN - 1545-2255
DOI - 10.1002/stc.2777
Subject(s) - waveform , convolutional neural network , hyperparameter , computer science , acceleration , feature (linguistics) , beam (structure) , artificial intelligence , fracture (geology) , pattern recognition (psychology) , structural engineering , engineering , physics , telecommunications , radar , linguistics , philosophy , geotechnical engineering , classical mechanics
Summary Beam‐end fractures of steel frame building structures cause the degradation of story stiffness, which may lead to the tilt or collapse of the whole structure. However, it is difficult to recognize this kind of local damage using conventional building damage detection methods. Generally, the beam‐end fracture can be recognized directly from the pulse of the waveform which is generated by the shock due to the fracture. Recently, with the development of technology of image classification and objects detection, the Convolutional Neural Networks (CNNs) have been proved as one of the effective methods for feature extraction. This study presents a beam‐end fracture detection method based on acceleration waveform using CNN model. 20,000 and 200 of pseudo acceleration waveform data which generated by numerical simulation were used to train and evaluate CNN model, respectively. To figure out the best hyperparameter configuration, grid search method was used in the hyperparameter optimization. One hundred percent accuracy is recorded in evaluation which is based on single‐degree‐of‐freedom (SDOF) database. Finally, the performance of trained and evaluated CNN model was verified by 80 real acceleration waveforms which are collected from two shake‐table tests. The verification result showed almost all the data was correctly classified except one nonfracture data was misclassified as the fracture data. The proposed method is sensitive to the sudden and drastical changes in terms of acceleration waveform and can correctly identify the feature of beam‐end fracture for real data with proper training. The accuracy of the proposed beam‐end fracture identification from the acceleration waveform was 98.75%.