To perform long-term structural health monitoring (SHM), a method based on a Nonlinear Autoregressive Exogenous (NARX) network is used to learn the features present in signals acquired from a pristine structure. When a subsequent measured signal is input to the trained NARX network, the output is a prediction of the equivalent signal from a pristine structure. The residual when the pristine predicted signal is subtracted from the measured signal is used for defect detection and localization. A methodology of how to train, test and assess a NARX network for guided wave signals is introduced and applied to experimental data obtained over a period of 8 years from a sparse array of guided wave sensors deployed on a steel storage tank. A separate NARX model is trained for each sensor pair in the array using data captured in 2012. The method is first tested using data from a single pair of sensors. Defect signals are synthesized by superposing simulated responses from defects onto later experimental signals obtained from the real structure. The test results for the NARX method show better detection performance than those from the Optimal Baseline Selection (OBS) method, in terms of Receiver Operating Characteristic (ROC) curves. The detection performance of NARX method is further assessed on signals from the whole sensor array, again with simulated defect responses superposed. It is shown that good detection and localization performance can be achieved by combining the NARX residual signals from different sensor pairs. The NARX method is tested on experimental data acquired at intervals over the following 7 years as the condition of the tank naturally degrades. Indications from localized corrosion are observed. Finally, an artificial localized anomaly is added to the tank and is visible at the correct location in the image formed using the NARX method.
|Journal||Structural Health Monitoring: An International Journal|
|Publication status||Published - 11 Jun 2021|
Bibliographical noteFunding Information:
The author(s) disclosed receipt of the following financial support for the research, authorship and/or publication of this article: The work reported is part of a pilot project (grant no. 100374) funded by Lloyd’s Register Foundation and the Alan Turing Institute Data-Centric Engineering Programme. Kangwei Wang was supported by the China Scholarship Council (grant no. 201906120066). Anthony J. Croxford was supported by an EPSRC fellowship (grant no. EP/M022528/1).
© The Author(s) 2021.