An effective framework for real-time structural damage detection using one-dimensional convolutional gated recurrent unit neural network and high performance computing

This paper proposes an effective one-dimensional convolutional gated recurrent unit neural network (1D-CGRU) by combining a one-dimensional convolutional neural network (1D-CNN) and a gated recurrent unit neural network (GRU) for real-time structural damage detection (SDD) based on time-series vibration signals measured from a large number of accelerators. In the proposed 1D-CGRU framework, the 1D-CNN is utilized for feature extraction in the spatial relation and for reducing the dimension of extracted feature vectors, while the GRU is used to learn the features in temporal relation and map extracted features to target outputs. In addition, to decrease the computational cost, a high performance computing (HPC) environment and a task parallelism process are used to train 1D-CGRU models for SDD using a large experimental dataset. The performance and efficiency of the proposed framework are investigated through a noise experiment dataset created by Qatar University (QU) Grandstand Simulator. The damage detection results achieved by the proposed approach are compared with the results attained by other approaches to verify the accuracy and reliability of the proposed approach. Moreover, the computing time of the parallel implementation in training 1D-CGRU models for SDD is compared with that of the sequential implementation to show the efficiency of the proposed task parallelism procedure. The obtained results have shown a great promise of the proposed framework for real-time and low-cost structural damage detection.