Damage Localization in Composite Plates Using Wavelet Transform and 2-D Convolutional Neural Networks

Abstract

Nondestructive evaluation of carbon fiber reinforced material structures has received special attention in the last decades. Usage of Ultrasonic Guided Waves (UGW), particularly Lamb waves, has become one of the most popular techniques for damage location, due to their sensitivity to defects, large range of inspection, and good propagation in several material types. However, extracting meaningful physical features from the response signals is challenging due to several factors, such as the multimodal nature of UGW, boundary conditions and the geometric shape of the structure, possible material anisotropies, and their environmental dependency. Neural networks (NN) are becoming a practical and accurate approach to analyzing the acquired data using data-driven methods. In this paper, a Convolutional-Neural-Network (CNN) is proposed to predict the distance-to-damage values from the signals corresponding to a transmitter-receiver path of transducers. The NN input is a 2D image (time-frequency) obtained as the Wavelet transform of the acquired experimental signals. The distances obtained with the NN are the input of a novel damage location algorithm which outputs a bidimensional image of the structure’s surface showing the estimated damage locations with a deviation of the actual position lower than 15 mm.

Diversion Detection in Small-Diameter HDPE Pipes Using Guided Waves and Deep Learning

“…The first step in the proposed DNN is to learn features from the data. Various methods have been used in the literature to extract or learn features using time and frequency domain features [23,25], using wavelets and a CNN [26,40], or using pre-established standard time-series classification networks, such as long short-term memory recurrent networks (LSTMs) [41]. In this paper, we use CNN and LSTM layers to learn features from the correlation envelope signals.…”

“…A recent review article [17] addressed the use of UGW and ML for defect detection. Some examples include structural health monitoring using piezoelectric sensors and ensemble learning by consensus using support vector machines [23] and using DNNs [24], defect sizing in pipes using ML algorithms and simulated and experimental data [25], estimating transducer-to-defect distance using the convolutional neural network (CNN) [26], assessing the degree of damage to a thin metal plate using wave-field measurements and DNN [27], and estimating the impact site of a steel ball on an aluminum plate using ML [28], among others. These works showed that high accuracy sensing is possible using their proposed methods.…”

Gain-Phase Error-Calibrated Piezoelectric Sensor Array-Based Impact Localization on Stiffened Curved Composite Structures

“…As known to all, multimodal and dispersive characteristics of Lamb waves cause the wave packet to overlap in the time domain and frequency domain, which challenges the signal interpretation. Azuara et al [ 6 ] pointed out that extracting meaningful physical features from the response signals is challenging due to the multimodal nature of Lamb waves and the geometric shape of the structure. Aircraft composite structures generally have stiffened components to enhance structural stiffness in practical applications.…”

Development of innovative techniques for lamb wave-based structural health monitoring systems in aeronautics

“… Distance-based damage localization algorithm, which introduces a new perspective based on distance-to-damage estimation, instead of using distances to estimate damage influence. This perspective allows using modern techniques, such as machine learning (ML) methods, to extract distance-based features from UGW signals [117].…”

“…Two different DL-based algorithms have been introduced: first, a regression-based CNN which outputs continuous distance values, and second, a classification-based CNN which outputs a discretized range of distance to damage and an additional evaluation of the damaged state of the inputted wavelet-transformed signal. The experimental evaluation of these solutions will be tested in section 6.5.2, and these methods were previously published in [117].…”