PSI - Issue 71

Rakesh Kumar Sahu et al. / Procedia Structural Integrity 71 (2025) 203–209

204

Nomenclature h (t)

Hanning window function

T h

Hanning window length

Central frequency of tone burst signal A Amplitude of signal In recent years, ultrasonic guided wave (Lamb wave)-based damage detection has emerged as a preferred technique over others due to its ability to propagate within structures with minimal attenuation and its high sensitivity to various structural defects, including voids, cracks, delamination, and discontinuities. The excitation and sensing of Lamb waves using a minimal number of sensors offers an efficient approach for monitoring large structural areas. Various algorithms have been developed for detecting, localization and severity assessment of damages in metallic and composite structures using Lamb waves (Kapuria and Agrahari, 2018), (Hua et al., 2020), (Zheng et al., 2024). Most studies have primarily focused on detecting single damage within structures, while the literature on multiple damage detection remains limited. To address the challenges associated with traditional multi damage detection methods, machine learning techniques have been increasingly utilized (Kim and Chattopadhyay, 2015; Tabian et al., 2019), with particular emphasis on neural networks (Rautela et al., 2021) and convolutional neural networks (Pandey et al., 2022). These methods are capable of processing complex signal networks to extract relevant features for detecting multiple defects. However, the accuracy and reliability of machine learning techniques are contingent on the quality of the training, testing, and validation processes (Pandey et al., 2022), as well as the size of the available data. Since raw sensor data contains extraneous information, it should not be directly input into the algorithm. Specific features, such as arrival time, mode, and amplitude, need to be extracted from the raw data based on the problem at hand. Therefore, large datasets are essential for accurate problem identification. Other machine learning algorithms, such as artificial neural networks (ANN) (Lopes et al., 2000) and deep neural networks (DNN) (Ewald et al., 2019; Pyle et al., 2021), are also available for addressing various challenges. In this study, damage severity is evaluated using damage-sensitive tone burst responses. The root mean square deviation (RMSD) damage index is determined by comparing the pristine state response with the response under damaged conditions for both single and multiple defects. Lamb wave interactions with defects, considering varying excitation frequencies, widths, and locations of damage, are simulated using the Abaqus/Explicit tool. To address the limitations of classical methods, a two-layer feed forward neural network is utilized to predict the width and location of multiple defects. For enhanced detection efficiency, a pure anti-symmetric mode response is extracted from the measured data, and damage-sensitive parameters such as the time of arrival of the residual peak, the number of peaks, and the standard deviation are used as inputs to the neural network. The organization of this paper is as follows: Section 2 provides methodology, damage index, and architecture of feed forward neural networks. Section 3 discussed the brief overview of numerical simulation and validation. Section 4 presents the results of the numerical simulation. While in Section 5, we discussed some important conclusions of the paper. The limitations, assumptions, and advantages are faithfully discussed. 2. Methodology Tone burst excitation: The Hanning-windowed narrowband tone burst excitation is the most commonly used excitation signal for damage detection using lamb waves to avoid the dispersion of the wave and the wave energy is concentrated at central frequency. The number of cycles and central frequency of the actuated signal should be optimized for selecting the smaller mode (Nandyala et al., 2020).

(1)

(2)