A framework of structural damage detection for civil structures using a combined multi-scale convolutional neural network and echo state network

Structural health monitoring (SHM) has become a notable method to ensure structural safety, yet the ability of existing damage detection techniques need improvements on extracting structural information from SHM data. Echo state networks (ESN) and multi-scale convolutional neural networks (MSCNN) proved effective in analyzing time and frequency domain data for civil structures. However, these models cannot identify structural information in the time-frequency domain. This study proposes a novel ESN-MSCNN combined model to effectively extract the time-frequency features of civil structures for damage detection. Firstly, vibration signal data is transformed into continuous time and Fourier spaces via data augmentation operation. Secondly, the ESN and MSCNN structures extract time and frequency domain features from preprocessed data, respectively. Finally, two combined features are fed into two fully connected layers to evaluate the degree of structural damage. Experiments on a scaled bridge and an IASC-ASCE benchmark building indicated that the proposed ESN-MSCNN model outperforms the state-of-the-art models for structural damage detection.