The localization of structural defects is of great interest in structure health monitoring (SHM).While acoustic emission signals are collected in the practice of SHM, the acquired waveformsinevitably include direct wave as well as reflection and reverberation waveforms. The directwave actually contains more straightforward information in localizing the sources, so in thiswork, a deep recurrent denoising autoencoder (DRDA) network is developed. In general,waveform signals are highly correlated at different timescales, so temporally recurrentconnections are added to the network structure, which have the memory of recent inputs.Consequently, the proposed DRDA model captures the dependencies across data points, whilecarrying out denoisng process, and combines the advantages of denoising autoencoders andrecurrent neural networks. As the output of the proposed DRDA, direct waveforms areextracted and validated through finite element simulations. A contrived structure with nontrivialshape is excited by simulated pencil break excitations under the ABAQUS environment,then the simulated responses provide training data for the DRDA. The proposed algorithm iseffective in filtering the reflected wave and outperforms the conventional denoising autoencoders.
展开▼