Defect detection in guided wave signals using nonlinear autoregressive exogenous method

Abstract

To perform long-term structural health monitoring, a method based on a nonlinear autoregressive exogenous 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 nonlinear autoregressive exogenous 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 nonlinear autoregressive exogenous 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 nonlinear autoregressive exogenous 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 nonlinear autoregressive exogenous method show better detection performance than those from the optimal baseline selection method, in terms of receiver operating characteristic curves. The detection performance of the nonlinear autoregressive exogenous 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 nonlinear autoregressive exogenous residual signals from different sensor pairs. The nonlinear autoregressive exogenous 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 nonlinear autoregressive exogenous method.

Enhancing damage localisation in critical areas of storage tanks in floating production storage and offloadings with a guided wave structural health monitoring system

“…[5][6][7][8] According to the used technique, diagnostic methods can be divided into: destructive technique when maximum load or fatigue test of samples are performed 9,10 ; and non-destructive technique when the state of the structure is evaluated without causing damage to it. 11,12 Due to this advantage, the nondestructive technique has become the most common technique for technical diagnostics nowadays.…”

“…For plate-like structures it can be applied through a sparse array of sensors in order to achieve coverage of large areas of the structure. [9][10][11] Inside a FPSO tank hull, different kinds of flaws can appear depending the position and corrosion protection method that is used. Uniform corrosion with thickness decrease and grooving corrosion are more likely to occur on plate surfaces and stiffener connections, respectively, whereas localised pitting appears more frequently in painted plate surfaces.…”

Potential and limitations of NARX for defect detection in guided wave signals

“…16 The data from the years 2012 and 2013, which are the immediate measurements after sensor installation, are assumed to represent the tank's pristine condition for the purposes of subsequent monitoring. The current work first investigated the NARX network in detail with data from one sensor pair: transmit on Sensor 1 and receive on Sensor The signals transmitted by the sensors are chirp excited, 22 and deconvolved to a five-cycle Hanning windowed toneburst with a centre frequency of 250 kHz, 16 and propagate primarily as S 0 mode. The data were sampled at 5 MHz, which gives 20 points per cycle.…”

“…The defect responses were scaled to the time of the first arrival signal, referred to as the Syn1 method in previous work, 16 to simulate the beam spreading effect, and were multiplied by another scale factor, referred to as severity (given in dB), 16 to represent the size of the defect response. Therefore, the amplitude a of the defect response is given by 16 : where c is the group velocity of the guided wave, t is the arrival time of the defect response and b is the severity of the defect. The current work initially looks at defects with severity of 230 dB occurring at t = 0.38 ms, to distinguish the detection performances of different networks using the largest signal to noise ratio (SNR) level not detected reliably.…”

“…This paper aims to improve the training robustness and physical interpretability of the network structure, by understanding how changing the length of history input, the width of the input layer and the number of future steps to predict ahead, affects its detection performance, using the same experimental data as previous work. 16 Subsequently, it aims to show the training procedure and detection performance of the network is robust to different initialisations of weights, defect responses at different times and data from different sensor pairs, revealing the potential of generalising such a network for similar defect detection problems in SHM and NDE.…”