Damage identification using the PZT impedance signals and residual learning algorithm

Damage identification techniques are of essential importance to promote the efficiency, reliability and safety of any structural system. In recent years, many artificial intelligence (AI)-based approaches have been successfully applied to establish damage identification tools using sample structural responses. However, it is generally difficult to fully train a deep neural network, therefore, researchers usually use shallow neural networks, which is limited in terms of performance. Addressing these issues, this paper proposes a novel structural damage identification method based on the raw time-series structural response signals and a deep residual network (DRN). A deep residual network is designed for extracting features of the raw time-domain impedance responses signals that measured from steel beam under different damage conditions. In order to optimize the network's performance, a residual learning algorithm and the Bayesian optimization algorithm are proposed and implemented. The results show that different structural conditions have been identified accurately. Also, the proposed methodology is suitable for processing structural responses signal with variable sequential length. Reasonable knowledge is required in damage detection and signal processing, which increases the applicability of the established method. Thus, the introduced method offers significant improvement for structural health monitoring (SHM) in terms of different damage sizes and location detection. To the best of our knowledge, this is the first work adopting DRN simultaneously on SHM non-image datasets of electro-mechanical impedance (EMI) signals.