Advertisement

JMST Advances

, Volume 1, Issue 1–2, pp 107–124 | Cite as

Damage assessment of smart composite structures via machine learning: a review

  • Asif Khan
  • Nayeon Kim
  • Jae Kyong Shin
  • Heung Soo KimEmail author
  • Byeng Dong Youn
Review
  • 352 Downloads

Abstract

Composite materials are heterogeneous in nature and suffer from complex non-linear modes of failure, such as delamination, matrix crack, fiber-breakage, and voids, among others. The early detection of damage in composite structures, such as airplanes, is imperative to avoid catastrophic failure and tragic consequences. This paper reports on the use of machine learning techniques for the damage assessment (i.e., detection, quantification, and localization) of smart composite structures. The success of the machine learning paradigm for damage assessment depends on the representational capability of the discriminative features for the problems of interest. However, from a practical standpoint, it is not possible to define a global or superset of discriminative features that could discriminate between damaged and undamaged states of the structures, and simultaneously make a distinction between various modes of failures. In addition, one machine learning algorithm may show optimum performance for the discriminative features of a particular problem but fails for others. This article focuses on a review of discriminative features and the corresponding machine learning algorithms (both supervised and unsupervised), for various types of damage in smart composite structures.

Keywords

Machine learning Composite materials Damage assessment Discriminative features 

1 Introduction

As the name indicates, composite materials refer to the types of materials that are obtained by combining two or more constituent materials of significantly different physical or chemical properties at a macroscopic level, without any solubility or chemical reaction, such that the resultant materials carry the best characteristics of its constituents. Currently, composite materials are obtained by reinforcing a matrix material with high-strength, high-modulus fibers, particles, or flacks. The matrix material can be either a polymer (epoxies, polyimides), metal (aluminum, titanium), or ceramic (alumina), whereas typical reinforcing materials are glass, graphite, silicon carbide, and boron. A laminated composite is a structural material that is formed by stacking together layers of fibrous composite materials to obtain tailored engineering properties, such as strength in specific directions, in-plane stiffness, bending stiffness, and coefficient of thermal expansion. Owing to their remarkable properties (e.g., high strength-to-weight ratio, tailored stiffness and strength, high fatigue life), composite materials are steadily replacing traditional materials in various industries, such as aerospace and aircraft, marine, wind energy, and automotive [1, 2], among others. However, due to their heterogeneous nature, composite materials are susceptible to various complex types of damage, such as delamination, fiber breakage, matrix crack, core disbonding, interlaminar porosity, and fiber misalignment [3]. The damages in composite materials may occur either during the manufacturing process or in service life and ultimately lead to the failure of composite structures. The initiation and progression of damages in composite structures follow a complex pattern, and cannot be detected with visual inspection and conventional techniques, because of the concealed and complex nature of damage e.g., delamination, matrix crack, and fiber breakage. Hence, the assessment of composite structures for the detection, quantification, and localization of damages is imperative for safe and reliable implementation of composite materials in real-world applications.

Materials that can respond to external stimuli (stress, temperature, electric field, magnetic field, moisture, pH, etc.) by altering one or more of their inherent properties in a controlled way are known as smart materials. For example, piezoelectric materials produce mechanical strain when an electric field is applied to them, and generate electric charges in response to mechanical deformation. The integration of smart materials (embedded or surface bonded, discrete or continuous) with structures in the form of sensor and actuator make them smart or intelligent structures. These smart structures are capable of sensing and adapting their static and dynamic response (adaptive structures), and continuously monitoring the presence, severity, and location of eminent damage from the sensor information [4, 5, 6, 7, 8]. Using the intelligent materials, a number of techniques have been developed in the last two decades for the damage assessment of smart composite structures, such as guided wave-based techniques [9, 10], wave field imaging-based techniques [11, 12], peridynamics based technique [13], and structural vibration based techniques [14, 15, 16, 17, 18, 19], among others.

From the literature, the most common domains for the extraction of discriminative features for the pristine and damaged composite structures are frequency domain, time domain, time–frequency domain, impedance domain, and modal analysis domain [20, 21, 22, 23, 24, 25]. The discriminative features in the aforementioned domains may be extracted either from low-frequency, high-wavelength structural vibration response, or high-frequency, low-wavelength guided waves or acoustic and ultrasonic waves [26, 27, 28, 29]. Some examples of discriminative features extracted from low-frequency, high-wavelength structural vibration response of composite structures are natural frequency, specific damping capacity, mode shape curvature, strain mode shapes, power spectral density, and shift in frequency response from a baseline [25, 30, 31, 32, 33, 34, 35]. Meanwhile, commonly employed descriptors or damage sensitive features for high-frequency, low-wavelength guided or acoustic waves are peak amplitude, rise time, duration, counts, counts to peak, energy, peak frequency, and average frequency [36, 37, 38, 39].

In recent years, machine learning techniques have emerged as powerful tools for the clustering, regression, and classification of discriminative features by drawing complex decision boundaries in the hyperplanes [40, 41, 42, 43]. However, in light of the preceding discussion, it is not possible to define a superset of discriminative features that could be used to distinguish between all kinds of damage modes (delamination, matrix crack, fiber fracture, voids, etc.) in composite structures, and simultaneously identify, quantify, and localize the particular types of damage with the same set of discriminative features. In addition, one machine learning strategy may successfully characterize one particular kind of damage, but fail for others. Therefore, the objective of this review article is to give a comprehensive review of the damage sensitive features for various types of damages in composite structures and the corresponding optimal machine learning tools, so that the reader can select the appropriate combination of discriminative features and machine learning tools for his or her problem of interest.

2 Fundamentals of machine learning

The main body of the article is dedicated to the review of damage sensitive features associated with various types of damage in smart composite structures, and the corresponding machine learning algorithms for those features. However, it will be helpful to briefly discuss the scope of machine learning. Machine learning is a scientific discipline that explores the study and development of mathematical algorithms that can construct functional relationships between quantities on the basis of known information and rules. According to Cherkassky and Mulier [44], machine learning techniques are employed to deal with the following three types of problem:
  • Classification, i.e., the development of a computational model by associating discrete labels with vectors of measured quantities. Once the model is developed from known data (training data), it can be used to predict labels for new measurements.

  • Regression, i.e., the development of a computational model on the basis of training samples, by associating continuous value targets or outputs with continuous value inputs.

  • Density estimation, i.e., the estimation of probability density functions from measured data samples with no output or target values.

In general, supervised learning techniques are used to deal with the problems of classification and regression, whereas density estimation is carried out via unsupervised learning body. Supervised learning is a body of knowledge that constructs computational models/rules on the basis of observational evidence. As the name suggests, supervised learning techniques learn the mapping function from the inputs to known outputs via learned and error-corrected approach, and then use those learned functions to predict output/labels for new unseen observations/inputs. Unsupervised machine learning techniques are concerned with determining the underlying pattern or structure within a data set of measurements. In contrast to supervised learning, unsupervised learning algorithms do not require any output or label for the observations. Figure 1 shows two main types of machine learning, their objectives, and some commonly used algorithms.
Fig. 1

Machine learning, its objectives, and commonly used algorithms

Worden et al. [45] have suggested seven axioms for structural health monitoring, from which the following three are particularly relevant to the current work.

Axiom III: Identifying the existence and location of damage can be done in an unsupervised learning mode, but identifying the type of damage present and the damage severity can generally only be done in a supervised learning mode.

Axiom IVa: Sensors cannot measure damage. Feature extraction through signal processing and statistical classification is necessary to convert sensor data into damage information.

Axiom IVb: Without intelligent feature extraction, the more sensitive a measurement is to damage, the more sensitive it is to changing operational and environmental conditions.

Hence, the extraction of damage sensitive features is of prime importance for the successful implementation of a machine learning paradigm for damage assessment. However, it is practically impossible to define a superset of damage sensitive features to assess and evaluate the conditions of all kinds of structural systems (e.g., civil, aerospace, oil and gas, etc.).

3 Damage assessment of laminated composites with machine learning

In this section of the article, we present a review of the damage assessment (i.e., detection, quantification, and localization) of composite structures via machine learning techniques, mainly focusing on the damage sensitive features and the machine learning tools for those damage sensitive features, so to give readers a general guideline on the type of discriminative feature and classification algorithm for those features. In the following subsections, we have reviewed the classification, clustering, and density estimation capabilities of artificial neural network (ANN), support vector machine (SVM), auto-regressive models (AR models), Bayesian classifier, K-means, and K-nearest neighbors on a variety of damage sensitive features in a supervised and unsupervised manner. The machine learning techniques that are developed from the combination of basic learning algorithms are reviewed under the section “miscellaneous”, and machine learning techniques that are not mentioned above are outlined in the section “others”. As discussed earlier, the main theme of this review article is to give a review of the damage sensitive features and machine learning algorithms for those features; therefore, we will not discuss the rigorous mathematical background of the learning algorithms.

3.1 Artificial neural network (ANN)

The history of ANN can be traced back to the late 1800 s, when scientific attempts were made to replicate the activity of the human brain. However, ANN was first employed for damage detection in 1991 by Ghaboussi et al. [46]. In literature, ANN has been successfully employed for the fault detection in induction motors [47], fault detection in bearing [48], gear fault detection and diagnostics [49], stochastic nonlinear system identification [50], and ship design [51], among others. In this section, ANN is reviewed for the damage assessment of composite structures. Islam and Craig [52] employed a back-propagation neural network for the quantification and localization of delamination in smart composite laminates. The neural network was trained with the first five modal frequencies as damage sensitive features, and the trained network was tested with unseen cases of delamination at different locations. Okafor et al. [53] used a feedforward backpropagation neural network to assess the delamination size in a smart composite beam. Non-dimensional delamination sizes and the corresponding first four modal frequencies were used to train the neural network. The trained network was tested with new cases of delamination using the first four normalized frequencies of test cases as input to the network. The network successfully predicted the dimensionless delamination size between (0.22 and 0.82), but could not predict the dimensionless delamination size below 0.08. Sung et al. [54] combined neural network with the Levenberg–Marquardt algorithm and the generalization method, for the accurate and reliable localization of low-velocity impact damage in smart composite laminates. The neural network was trained, cross-validated, and tested with the differential arrival time of impact-generated acoustic waves to PZT sensors as input to the network, and the location of the impact as target output. For a graphite/epoxy laminate of (330 mm × 330 mm × 3 mm), the proposed technique predicted the location of impact under the error of 5 mm. Sammons et al. [55] combined x-ray computed tomography with a convolutional neural network (CNN), for the identification and quantification of delamination in carbon fiber reinforced polymer composite. A class label was assigned to each patch (201 × 201 pixels) of the image based on the central pixel in the corresponding patch. The trained network was able to identify and quantify delaminations of small sizes, but could not approximate the size of large delaminations. Bar et al. [56] employed artificial neural network (ANN) for the identification and classification of failure modes in glass fiber reinforced plastics composites from the AE signals of static tensile tests measured with a surface mounted PVDF film. The Kohonen self-organizing feature map (KSOM) was employed to generate class information for different modes of failure (i.e., matrix crack, fiber-matrix debonding, fiber fracture). The class information of KSOM was used for the training of the multilayer perceptron (MLP). The proposed technique successfully discriminated different failure modes, irrespective of the overlapping in the parameters of AE signals associated with those modes. Bhat et al. [57] employed ANN to eliminate the noise of various types from AE signals and characterize AE events belonging to three different failure modes in CFRP specimens. The attributes of the AE single being used as descriptor were rise time, ring down counts, energy, and peak amplitude. It was shown that KSOM and MLP were able to successfully separate and classify noise of different sources from failure modes (fiber failure, fiber/matrix debond, matrix crack) of the CFRP. Table 1 summarizes the classification results of ANN on the training data of noise (N) and various failure modes [F1 (fiber failure), F2 (fiber/matrix debond), and F3–F6 (matrix related failures)].
Table 1

Training results for seven-class classification [57]

Class

No. of signals present

No. of signals classified into that class

Break-up details

% Correct classification

N

F1

F2

F3

F4

F5

F6

N

1000

993

911

11

1

70

91.74

F1

1000

987

984

3

99.69

F2

1000

961

1

940

19

1

97.81

F3

1000

1013

16

997

98.42

F4

1000

982

3

978

1

99.59

F5

1000

1044

3

60

4

977

93.58

F6

1000

1020

82

7

3

928

90.98

Chetwynd et al. [58] employed multilayer perceptron neural network for the classification and regression problems of damage detection in a stiffened curved carbon fiber reinforced panel with surface bonded piezoelectric transducers. A number of localized damages were simulated via a force applicator, and Lamb wave responses were obtained for the healthy and damaged cases. The Lamb wave response for each case was converted into a scalar novelty index via outlier analysis. The novelty indices of 28 sensor paths (ignoring reversed paths) of eight surface-mounted sensors were used as input to the MLP classification and regression networks. The MLP classification network was used for the classification of damages and undamaged regions of the panel, whereas the MLP regression network was employed for estimating the exact location of damage on the panel. Table 2 shows the classification performance of MLP classification network in terms of confusion matrix, whereas, Fig. 2 shows the performance of a trained MLP regression network for predicting the locations of damages.
Table 2

Confusion matrix for third-level Daubechies-4 MLP classification network [52]

True class

Predicted class

Undamaged

Damage region 1

Damage region 2

Damage region 3

Undamaged

40

0

0

0

Damage region 1

1

189

7

3

Damage region 2

0

10

88

2

Damage region 3

4

15

7

174

Fig. 2

Average MLP test-set predictions and true damage locations [52]

From Table 2, the MLP classification network showed an overall classification accuracy of 90.9%.

3.2 Support vector machine (SVM)

Support vector machine (SVM) maps the linearly inseparable feature space to higher dimensional space with the help of kernel function, such that a maximal separating linear hyperplane can be constructed in the higher dimensional space. The mathematical details of SVM can be found in Refs. [59, 60, 61]. In the literature, SVM has been used for fault diagnosis of rotating machine [62], gear fault diagnosis under variable conditions [63], source localization of acoustic emission [64], and defect diagnostics of gas turbine engine [65, 66], among others. Das et al. [67] employed one-class support vector mechanics (SVMs) to identify and classify four categories (notches, saw cut, drilled holes, delamination) of damage in smart composite laminates. They used a time–frequency-based approach (Gabor’s spectrogram technique) and a time-embedding technique (tapped-delay approach), to extract damage sensitive features from the response of piezoelectric sensors. The classification results of one-class SVMs revealed that time-embedded features are better than time–frequency features for classification accuracy. Prashant and Sung [68] developed an SVM-based methodology for online damage detection in composite helicopter rotor blade. The vibratory hub loads and moments of the stiff-in-plane hingeless composite helicopter rotor blade were used as damage sensitive features for matrix cracking and debonding/delamination damage. The damage sensitive features were classified into three classes (matrix cracking up to saturation, initiation of delamination, and severe delamination) via SVM. Farooq et al. [69] studied the performance of SVM and ANN for the detection and severity assessment of crack damage in fiber composite panel. Static strain mixed with Gaussian noise was used for training and cross-validation of SVM and ANN. SVM was found to perform better than ANN for damage detection and severity assessment. Dib et al. [70] developed a novelty classifier on the basis of one class SVM for damage detection. The performance of the classifier was evaluated on the data from impact damage experiments on a glass fiber composite plate. For discriminate features, the guided wave was segmented into L time bins and Fourier transform was computed for each time bin. The feature vector was constructed from the amplitude and phase of the signal in each time bin.

3.3 Auto regressive models (AR models)

Auto-regressive (AR) model is a statistical tool that is used to describe time-varying processes by forecasting a variable of interest in terms of the previous values of the same variables and parameters of the AR model. Some relevant examples of AR models for damage detection are fault diagnostic and condition monitoring of rolling bearing [71], damage detection in wind turbine [72], damage detection in frame structures [73], and structural health monitoring of civil infrastructure [74, 75]. Nardi et al. [76] studied the detection of low-velocity impact-induced delamination in smart composite laminates via auto-regressive (AR) models. An optimal order was identified for the AR model to fit the undamaged/damaged data, and the corresponding AR parameters were transformed into a new sub-space via linear discriminant analysis (LDA). The within-class (Sw) and between-class (Sb) scatter matrices of the transformed AR parameters were used to define a new matrix \(V\;(V = S_{\text{w}}^{ - 1} \times S_{\text{b}} )\). The first three eigenvectors and the corresponding eigenvalues of the matrix V were chosen as damaged sensitive features. Figure 3 shows the clusters of different damaged and undamaged cases formed for V on the basis of the proposed technique.
Fig. 3

Reduced dataset in the reduced state space: C1U (asterisk), C1D (cross), C2U (square), C2D (circle), C3U (diamond), C3D (triangle). AR model order p = 25 [76]

Herein, the undamaged states of the three plates are shown by C1U (asterisk), C2U (square) and C3U (diamond), whereas, damaged states of the same plates are shown by C1D (cross), C2D (circle) and C3D (triangle). The C1D, C2D and C3D were, respectively, subjected to an impact energy of 20 J, 8 J and 12 J.

Vamvoudakis-Stefanou et al. [77] compared the performance of two non-parametric methods [i.e., power spectral density (PSA) and transmittance function (TF)] and one parametric method [autoregressive model (AR)], for the detection of impact damage in composite beams. All the three methods were trained with random vibration response signals of a healthy beam. The effectiveness of the trained models for the damage detection was evaluated by testing those with the random vibration responses from 22 healthy and 3 damaged beams. It was found that among the three methods for damage detection, the parametric AR based method is more effective and robust.

3.4 Principle component analysis (PCA)

In general, principal component analysis (PCA) has been extensively employed to reduce the dimensionality of damage sensitive features or structural dynamic response signals [78, 79]. In some cases, dimensionality reduction allows for visualizing the clusters of damage sensitive features and provides interesting results on the detection and classification of damages [26]. Other applications of PCA are sensor validation [80], and distinguishing of the environmental effects from structural damage [81, 82], among others. Oskouei et al. [83] combined fuzzy c-means (FCM) clustering with principal component analysis for unsupervised clustering of acoustic emission data, with the aim of discriminating the acoustic signatures of damage mechanisms in glass/polyester composite materials. The rise-time, counts, energy, amplitude, peak frequency and duration of the AE signals were used as descriptors for matrix cracking, fiber-matrix debonding, and fiber failure. Figure 4 shows the PCA visualization of the FCM clustering. Herein, three classes can be identified for each of the test specimens. Frequency distribution of each cluster was employed to discriminate the clusters corresponding to matrix cracking, fiber-matrix debonding, and fiber failure, as shown in Fig. 5.
Fig. 4

PCA visualization of the fuzzy c-means clustering: a T3, b T4, and c T5 [83]

Fig. 5

Frequency distribution of the patterns belonging to each cluster at different interfaces. a T3, b T4, and c T5. C1: matrix cracking, C2: fiber debonding, C3: fiber breakage [83]

3.5 Bayesian classifiers

Bayesian classifiers work on the basis of the Bayesian theory, which is a statistical tool that makes use of probabilities and costs for quantifying the tradeoff between various decisions. A learning body that works on the Bayesian decision theory employs the concepts of Bayesian statistics to estimate the expected values of its actions, and to update its expectations based on new information [84, 85, 86]. Peng et al. [87] proposed a probabilistic framework for the quantification and localization delamination in smart composite laminates. Correlation coefficient and phase change of the Lamb wave-based signal were employed as damage sensitive features. The damage size and location were estimated by incorporating the damage sensitive features into the Bayesian updating framework. Nguyen et al. [88] studied the feasibility of Bayesian network for the detection of hole-type damages on a smart composite plate. The difference between the amplitudes and the time of occurrence of the 1st and 2nd peaks of the fundamental Lamb wave modes of the sensor paths were chosen as damage sensitive. The parameters used for the training of Bayesian network were the location (x, y) of the hole, area (a, b) of the hole, and two damage sensitive features from the sensor paths. The trained network predicted the location and area of the hole in a probabilistic manner. Rabiei et al. [89] studied an SHM framework for the monitoring and prognostics of complex multi-stage degradation process in a composite material under fatigue, by employing dynamic Bayesian network, along with extended particle filtering and support vector regression. Dissipated thermal energy, temperature, and AE counts were employed as damage indices. Addin et al. [90] studied the performance of Naïve-Bayes for the classification of three types of damage (i.e., delamination, crack, and hole) in quasi-isotropic graphite/epoxy laminated composites. The damage sensitive features were based on Lamb waves that were produced and received via surface bonded PZT actuators and sensors. A feature sub-set selection method based on k-means algorithm was employed to select the optimum number of features from the amplitude values of the Lamb waves. Nichols et al. [91] proposed a Bayesian framework to approximate the non-linear parameters of a system from the free-decay vibrations. The Markov Chain Monte Carlo (MCMC) approach was employed to estimate the Bayesian parameters of a non-linear system from free-vibration, time-series data. The proposed approach was applied for the damage assessment (location, size, and depth of delamination) of a cross-ply carbon epoxy laminate. Figure 6 shows the performance of the proposed approach for estimating the length of delaminations by approximation of the end-points of delaminations.
Fig. 6

Estimated delamination end point as a function of the actual delamination end point. The intervals of confidence were created as the central 95 percent of the sampled values from p(xb) obtained using the MCMC algorithm [85]

3.6 K-means and K-nearest neighbors

K-means is an unsupervised learning technique that finds K-groups or K-clusters within an unlabeled data set. The term K specifies the number of groups or clusters. The K-means algorithm assigns each data point within the data set to a particular group, based on the distance of the point from the center of the given group. On the other hand, K-Nearest Neighbors (KNN) is a supervised learning technique that is commonly employed for clustering and classification. The ‘K’ in KNN specifies the number of nearest neighbors used to classify a test sample. Godin et al. [92] employed supervised and unsupervised classifiers for clustering of the acoustic emission signals from the tensile tests of unidirectional glass/polyester composite. AE signals were first clustered via k-means on the basis of rise-time, the number of counts, duration, amplitude, and counts to the peak of the AE signals. The clusters identified with k-means were used as labeled data for the supervised classifier of KNN. The trained KNN classifier was employed for the classification of new data. The unsupervised classifier of Kohonen’s map showed the same results as the KNN classifier. The proposed approach successfully classified the AE signal associated with matrix cracking and interfacial decohesion. Pashmforoush et al. [93] combined k-means algorithm with genetic algorithm for the clustering of various damage mechanisms in a mode I delamination test of a sandwich composite. Figure 7 shows the employed procedure for unsupervised damage classification.
Fig. 7

A concise explanation of the damage classification procedure [93]

Herein, a mode I delamination test was conducted on a sandwich composite material and its key failure events (fiber breakage, matrix crack, and core damage) were captured with an AE sensor. The discriminative features of AE signals were identified for each failure mode and principal component analysis (PCA) was employed to reduce the dimension of the discriminative feature space. The reduced dimensioned features were clustered into different groups through an integration of genetic algorithm and k-means algorithm. The identified clusters were assessed for the distinct damage mechanism. The results were validated through microscopic observations by SEM.

This study showed that frequency (Fig. 8) was the best descriptor among different AE parameters for different damage mechanisms i.e., fiber breakage, adhesion failure, core failure, and matrix crack.
Fig. 8

Frequency—amplitude plot of clustered AE signals in specimen S1 [93]

3.7 Miscellaneous

This section of the article provides a review of the machine learning frameworks that are developed from the combination of basic machine learning tools (e.g., ANN, PCA, KNN, LDA). The main purpose of using more than one learning algorithms could be one of the following: enhance the discriminative capabilities of the damage features, extract discriminative features, reduce the dimensions of the discriminative features, identify the hidden pattern in the discriminate feature space, or to select the most appropriate discriminate features. The word ensemble classifiers have been used in the literature for a combination of classifier whose final predictive decision is a combination (typically by weighted or unweighted voting) of all the classifiers. More details on ensemble classifiers can be found in the references [94, 95]. Gaudenzi et al. [96] developed an experimental technique for the classification of delamination damage in smart composite laminates. The delamination was induced by low-velocity impacts of various energies. They employed wavelet packet transform (WPT) to extract wavelet-based damage sensitive features from the dynamic response of the damaged and undamaged smart plates. The discriminative capability of the extracted features was enhanced via linear discriminant analysis (LDA). The damaged and undamaged cases were classified in a supervised manner through the normalized Euclidean distance between the damage sensitive features of intact and damaged samples. McCrory et al. [97] employed artificial neural network (ANN), unsupervised waveform clustering (UWC), and measured amplitude ratio (MAR) to characterize acoustic emission (AE), and to subsequently classify the nature of damage in a carbon fiber composite panel under a buckling regime. The ANN was trained with the rise-time, counts, absolute energy, duration, amplitude, average frequency, central frequency, and peak frequency of AE. For UWC, instead of the complete waveform, a normalized segment of the AE waves was used as input to PCA, to extract uncorrelated features based on waveform appearance. The choice of segments limited the dependency of the clustering process on the varying parameters of AE over the propagation distance from the damage source to the sensor. MAR was employed to distinguish between the delamination and matrix cracking, based on the ratio of flexural (A0) and extensional (S0) modes. Before calculating the MAR of each recorded signal, both the A0 and S0 modes were corrected for the possible attenuations. ANN and UWC identified two clusters, and MAR values were used to distinguish between those clusters as delamination (MAR < 1) and matrix cracking (MAR > 1). Kessler and Rani [98] coupled Lamb wave testing with PCA and pattern recognition to predict the presence, type, and severity of damage in smart composite laminates. PCA was used to extract damage sensitive features from the Lamb wave response of composite plates with three types of damage (impact, hole, and delamination). The first 20 principal components were selected as damage sensitive features and used those features to compare the performance of KNN, ANN, and Decision Tree for predicting the presence, type, and severity of damage. The optimized KNN algorithm was found to be the most reliable method for the detection, classification, and evaluation of damages in smart composite laminates. Oliveira and Marques [99] proposed a procedure for the identification and discrimination of damage in smart composite materials by clustering acoustic emission signals with artificial neural networks. Table 3 summarizes the time and frequency domain features that were extracted from the AE signals. The discriminative features were first clustered by a self-organizing map of Kohonen (SOM), and then k-means was employed to cluster the SOM. Figure 9 shows the clustering results. For the 10 clusters of Fig. 9, six AE waveforms were identified from the modal nature of the AE signal that corresponded to damage sequence in the smart structure.
Table 3

Selected features of the AE signals [99]

Selected features

Unit

Amplitude

mV

Rise time

μs

Duration

μs

Energy

1e−7 V2s

Counts

Counts to peak

Amplitude ratio 1

Amplitude ratio 2

E/F

Counts/duration

kHz

FFT peak frequency 1

kHz

FFT peak frequency 2

kHz

FFT amplitude

Spectrum’s centre of gravity

kHz

Fig. 9

Map clustering [99]

Li et al. [100] carried out cluster analysis of AE signals that originated from the damage initiation and development of 2D and 3D glass/epoxy woven composites. Peak amplitude (PA) and peak frequency (PF) of the AE signature were identified as the most effective discriminative features. The AE events were clustered through k-means++ clustering algorithm and PCA. Silhouette coefficient and Davies–Bouldin index were employed to identify the optimum number of clusters. Figure 10 shows the optimum number of clusters for the specimens of plain weave composite in wrap or fill direction (PW), 3D woven composite in wrap direction (3W) and 3D woven composite in fill direction (3F), whereas, Table 4 shows the bounds of each cluster.
Fig. 10

Cluster results separated by PA and PF for all specimens: a PW-1, -2, -3; b 3W-1, -2, -3; c 3F-1, -2, -3 [92]

Table 4

Cluster bounds of PW and 3D textile composites [92]

Cluster bounds

PA (dB)

PF (kHz)

CL1

35–55

50–80

CL2

35–55

80–150

CL3

55–100

50–150

CL4

35–80

150–500

Gutkin et al. [39] analyzed AE signals with unsupervised pattern recognition algorithms [k-means, self-organizing map (SOM) combined with k-means, competitive neural network (CNN)], and peak frequency content for the failure investigation in CFRP. SOM combined with k-means appeared to be the most effective in clustering of different failure modes, on the basis of peak amplitude, peak frequency, energy, rise time, and duration of the AE signals. The peak frequency content of each test revealed different modes of failure, as shown in Fig. 11.
Fig. 11

Summary of the classification using frequency content [39]

Godin et al. [101] proposed the combination of KSOM and k-means for the classification of AE signals and monitoring the chronology of damage events in cross-ply glass/epoxy composites. They employed amplitude, duration, rise time, counts, counts to peak, and energy of the AE signals as the descriptors for matrix cracking, interfacial debonding, and fiber failure of the composites. Pashmforoush et al. [102] proposed the integration of harmony search and k-means algorithm for acoustic emission-based damage classification of glass/polyester composites of three different lay-up configurations. The AE parameters of rise-time, peak amplitude, energy, count, average frequency, and duration were employed as discriminative features. The frequency was identified as an efficient descriptor for damage characterization. It was found that AE signals of highest and lowest frequency contents were representative of fiber breakage and matrix cracking, respectively, while the debonding occurs in a frequency range that was between the fiber breakage and matrix cracking. Scanning electron microscopic (SEM) observations were used to verify the classification results. Al-Jumaili et al. [103] studied k-means and Fuzzy C-means techniques for unsupervised classification of acoustic emissions from a carbon fiber laminate buckling test. A hierarchical clustering algorithm was used to evaluate the correlation of 17 AE parameters i.e., rise time, counts, energy, duration, amplitude, average frequency, absolute energy, centroid frequency, peak frequency, decay time, rise time/duration, amplitude/rise time, amplitude/decay time, amplitude/average frequency, absolute energy/amplitude, duration/amplitude and rise time/decay time. PCA was employed to reduce the dimensionality of the features selected via hierarchical clustering. The unsupervised techniques of k-means and Fuzzy C-means successfully classified the source mechanisms (matrix crack, delamination) into separated clusters. Furthermore, Delta T Mapping method correctly identified the damage locations. Tibaduiza et al. [104] studied multiway principal component analysis (MPCA), discrete wavelet transform (DWT), squared prediction error (SPE) and self-organizing maps (SOM) for the detection and classification of damages in an aircraft fuselage and a CFRP composite plate. Damage sensitive features were extracted and selected via PCA and SPE, respectively, either directly from the dynamic responses, or the wavelet coefficients. The damage sensitive features were clustered via SOM algorithm to detect and isolate the outliers that refer to the different structural states in different zones. Figures 12 and 13 show the classification of undamaged (UN) and damaged states (D1, D2, D3, D4, D5, D6) of the aircraft fuselage and composite plate, respectively, using 3 scores and SPE.
Fig. 12

Damage classification using 3 scores and SPE in the aircraft fuselage. a, d Shows the cluster map and U-matrix using raw time signals. b, e Shows the cluster map and U-matrix using the approximation coefficients. c, f Shows the cluster map and U-matrix using the detail coefficients [96]

Fig. 13

Damage classification using 3 scores and SPE in the composite plate. a, d Shows the cluster map and U-matrix using raw time signals. b, e Shows the cluster map and U-matrix using the approximation coefficients. c, f Shows the cluster map and U-matrix using the detail coefficients [96]

Vitola et al. [105] employed multivariate analysis, sensor data fusion, and machine learning to develop an SHM framework for structures subjected to temperature variations. The methodology was implemented for the damage assessment of aluminum and composite structures. PCA was used to form a discriminative feature space. Table 5 shows the performance of twenty different supervised machine learning techniques for damage identification in a composite plate. Herein, the columns correspond to the percentage of correctly classified instances for the healthy and damaged cases.
Table 5

Percentage of correct decisions for the healthy structure and the structure with damages 1, 2, and 3, for the twenty different machine learning strategies (composite plate) [105]

Machine Name

Healthy (%)

Damage 1 (%)

Damage 2 (%)

Damage 3 (%)

Medium tree

55

63.33

60.83

52.5

Simple tree

40

60

63.33

42.5

Complex tree

57.5

64.17

75.83

65.83

Linear SVM

41.67

59.17

45

47.5

Quadratic SVM

65.83

73.33

85

75.5

Cubic SVM

70.83

75

86.67

74.17

Fine Gaussian SVM

59.17

64.17

83.33

78.33

Medium Gaussian SVM

55.83

60

82.5

63.33

Coarse Gaussian SVM

52.5

10.83

33.33

56.67

Fine k-NN

63.33

61.67

80

70

Medium k-NN

65

46.67

75

63.33

Coarse k-NN

52.5

37.5

60.83

35.83

Cosine k-NN

65

43.33

79.17

60.83

Cubic k-NN

59.17

47.5

72.5

60

Weighted k-NN

61.67

58.33

83.33

74.17

Boosted trees

16.67

62.5

60.83

71.67

Bagged trees

71.67

72.5

90

84.17

Subspace discriminant

33.33

45.83

45

55.83

Subspace k-NN

70.83

72.5

89.17

82.5

Rusboosted trees

0

62.5

0

93.33

Recently, Jimenez et al. [106] studied the classification algorithms of quadratic discriminant analysis, KNN, decision trees, and MLP for the assessment of delamination damage in wind turbine blades. The damage sensitive features were extracted from the wavelet transform of guided waves via autoregressive Yule-Walker model, and Akaike’s information criterion was employed to select the most appropriate features for the machine learning algorithms.

3.8 Others

This section of the article provides an overview of those learning algorithms that are either not very common, or newly developed. Tang et al. [107] proposed a statistical unsupervised approach to distinguish the changes in guided wave signals due to the combined effects of temperature, load, and vibration from the changes associated with structural damage (crack in a smart aluminum plate, impact damage in a smart composite laminate). The sensed guided wave signals were processed via wavelet transform, and wavelet energy was selected as a damage sensitive feature. Monte Carlo procedure was employed to estimate the probability density functions (PDFs) of the damage sensitive feature of the healthy and damaged coupons. The PDFs were compared to differentiate the presence of damage from environmental effects in a probabilistic manner, as shown in Fig. 14.
Fig. 14

Fitted normal PDFs of wavelet energy data for composite coupon before and after damage a 200 kHz, and b 300 kHz [107]

Sohn et al. [108] proposed an unsupervised damage classifier for the identification and localization of delamination area in a piezo-bonded quasi-isotropic graphite/epoxy composite plate. A wavelet analysis procedure was employed to extract damage sensitive features. A damage index was defined on the basis of the ratio of the kinetic energies of the time response at the input frequency of the test signal and baseline signal. Extreme value statistics was employed to compute a damage threshold value, which was used to classify the responses from damaged or undamaged structures. They also identified the location and size of delamination via virtual grids. Sultan et al. [109] employed the concept of discordancy from the statistical discipline of outlier analysis, to identify and quantify the damaging and non-damaging impact in smart composite laminates. Frequency centroids were obtained from the transient response of surface-mounted PZT sensors, and those centroids were employed as damage sensitive features. The frequency centroid of each individual sensor was treated as a univariate damage feature, whereas combining the frequency centroids of all sensors for each case formed a multivariate damage sensitive feature. To detect outliers, a deviation statistic based discordancy test was used for a univariate data set, and Mahalanobis squared-distance was employed for the multivariate data. The data sets of the univariate and multivariate damage sensitive features were clearly separated into damaging and non-damaging impacts by a threshold value, as shown in Figs. 15 and 16.
Fig. 15

Damage indices for non-damaged and damaged cases; (solid line) threshold value: a sensor 1, b sensor 2, and c sensor 3 [109]

Fig. 16

Mahalanobis squared distances for non-damaged and damaged cases; (solid line) threshold value [109]

Fotouhi et al. [110] studied the relationship between AE signals of quasi-static three-point bending test and damage mechanisms of woven [0, 90]s and unidirectional [0]s glass/epoxy composites, by employing Fuzzy C-means (FCM) clustering and wavelet packet transform (WPT). Peak amplitude, counts, and average frequency of the acoustic signal were used as descriptors for different fracture mechanisms (matrix cracking, debonding, and fiber breakage). It was found that FCM successfully classified the AE for the three damage modes, and that the fracture mechanisms for the woven (A1) and unidirectional (A2) samples were different, as shown in Tables 6 and 7.
Table 6

Central characteristic of three classes obtained from FCM classification for specimen A1 [110]

Signal parameters

Average frequency (kHz)

Peak amplitude (dB)

Count

First class

139.95

49.23

10.18

Second class

265.51

53.47

12.61

Third class

412.2

55.12

11.74

Table 7

Average dependency percentage of signals for three classes [110]

Experimental condition

Dependency on first class

Dependency on second class

dependency on third Class

Specimen A1

54%

37%

9%

Specimen A2

11%

24%

65%

4 Conclusion

The continuously increasing use of composite materials in various industries, such as aerospace, automotive, and wind energy, among others, demands early assessment (detection, localization, and quantification) of damage in these materials for the enhancement of their structural integrity and reliability. The emergence of machine learning for self-driving cars, smart homes, automated surveillance systems, and natural language processing has caught the attention of the research community. In this review article, supervised and unsupervised machine learning techniques have been reviewed for the damage assessment of smart composite structures, with special emphasis on the types of damage sensitive features for various kinds of damages. The following is a summary of the reviewed articles.
  1. 1.
    The most common types of discriminative features from the damaged and undamaged states of the composites for the training, cross-validation, and testing of learning algorithms are:
    • Natural frequencies [for (5–10) fundamental modes], wavelet energy packets, vibration hub loads, auto-regressive parameters, and wavelet coefficients from low-frequency structural vibration responses.

    • Arrival time, rise time, ring down counts, energy, peak amplitudes, average frequency, central frequency, peak frequency, duration, amplitude, and counts to peak from acoustic emission signals.

    • Correlation coefficients, phase change, difference of amplitudes and time of occurrence of fundamental modes, amplitudes, wavelet energy, and frequency centroids from guided Lamb waves.

    • Number of Pixels for image-processing based damage assessment via deep learning.

     
  2. 2.

    Unsupervised learning may be employed to generate class information for different modes of failures, and the labeled data can then be used for the development of predictive models.

     
  3. 3.

    Frequency distribution may be employed to associate the clusters of unsupervised methods with the corresponding modes of failure.

     
  4. 4.

    Acoustic emission signals of highest and lowest frequency contents are representative of fiber breakage and matrix cracking, respectively, while debonding occurs in a frequency range that is between those of fiber breakage and matrix cracking.

     

In general, conventional machine learning algorithms (e.g., SVM, decision trees) are computationally less expensive, depends on low-end machines (CPUs), perform well on small learning data and require handcrafted features from the source data, whereas, deep learning algorithms (e.g., CNN) are computationally more expensive, depend on high-end machines (GPUs), require large learning data, and automatically extract discriminative features from the raw input data (e.g., time/frequency domain signals).

Notes

Acknowledgements

This research was supported by the National Research Council of Science & Technology (NST) grant by the Korea Government (MSIT) (CAP-17-04-KRISS) and was also supported by the Basic Science Research Program through the National Research Foundation of Korea (NRF-2017R1D1A1B03028368), funded by the Ministry of Education.

References

  1. 1.
    A. Elmarakbi, Advanced composite materials for automotive applications: Structural integrity and crashworthiness (Wiley, Chichester, 2013)Google Scholar
  2. 2.
    D. Gay, Composite materials: design and applications (CRC Press, Boca Raton, 2014)Google Scholar
  3. 3.
    R. Smith, Composite defects and their detection. Mater. Sci. Eng. 3, 103–143 (2009)Google Scholar
  4. 4.
    G.H. Park, C.R. Farrar, A.C. Rutherford et al., Piezo-sensor self-diagnostics using electrical impedance measurements (Los Alamos, Los Alamos National Laboratory, 2004)Google Scholar
  5. 5.
    J.W. Sohn, H.S. Kim, Active recovery of vibration characteristics for delaminated composite structure using piezoelectric actuator. Int. J. Precis. Eng. Manuf. 16(3), 597–602 (2015)Google Scholar
  6. 6.
    A. Khan, H.S. Kim, Active vibration control of a piezo-bonded laminated composite in the presence of sensor partial debonding and structural delaminations. Sensors 19(3), 540 (2019)Google Scholar
  7. 7.
    A. Khan, H.S. Kim, Assessment of sensor debonding failure in system identification of smart composite laminates. NDT E Int. 93, 24–33 (2018)Google Scholar
  8. 8.
    A. Khan, H.S. Kim, B.D. Youn, Modeling and assessment of partially debonded piezoelectric sensor in smart composite laminates. Int. J. Mech. Sci. 131, 26–36 (2017)Google Scholar
  9. 9.
    A. Raghavan, Guided-wave structural health monitoring, 2007Google Scholar
  10. 10.
    Z. Su, L. Ye, Y. Lu, Guided Lamb waves for identification of damage in composite structures: a review. J. Sound Vib. 295(3–5), 753–780 (2006)Google Scholar
  11. 11.
    M.D. Rogge, C.A. Leckey, Characterization of impact damage in composite laminates using guided wavefield imaging and local wavenumber domain analysis. Ultrasonics 53(7), 1217–1226 (2013)Google Scholar
  12. 12.
    H. Sohn, D. Dutta, J. Yang et al., Delamination detection in composites through guided wave field image processing. Compos. Sci. Technol. 71(9), 1250–1256 (2011)Google Scholar
  13. 13.
    A. Yaghoobi, M.G. Chorzepa, Meshless modeling framework for fiber reinforced concrete structures. Comput. Struct. 161, 43–54 (2015)Google Scholar
  14. 14.
    D. Montalvao, N.M.M. Maia, A.M.R. Ribeiro, A review of vibration-based structural health monitoring with special emphasis on composite materials. Shock Vib. Dig. 38(4), 295–324 (2006)Google Scholar
  15. 15.
    H.S. Kim, A. Ghoshal, J. Kim et al., Transient analysis of delaminated smart composite structures by incorporating the Fermi–Dirac distribution function. Smart Mater. Struct. 15(2), 221 (2006)Google Scholar
  16. 16.
    H.S. Kim, A. Ghoshal, A. Chattopadhyay et al., Development of embedded sensor models in composite laminates for structural health monitoring. J. Reinf. Plast. Compos. 23(11), 1207–1240 (2004)Google Scholar
  17. 17.
    A. Khan, D.-K. Ko, S.C. Lim et al., Structural vibration-based classification and prediction of delamination in smart composite laminates using deep learning neural network. Compos. B Eng. 161, 586–594 (2019)Google Scholar
  18. 18.
    A. Khan, H.S. Kim, Assessment of delaminated smart composite laminates via system identification and supervised learning. Compos. Struct. 206, 354–362 (2018)Google Scholar
  19. 19.
    A. Khan, H.S. Lee, H.S. Kim, Analysis of sensor-debonding failure in active vibration control of smart composite plate. J. Intell. Mater. Syst. Struct. 28(18), 2603–2616 (2017)Google Scholar
  20. 20.
    W. Staszewski, C. Boller, G.R. Tomlinson, Health monitoring of aerospace structures: smart sensor technologies and signal processing (Wiley, New York, 2004)Google Scholar
  21. 21.
    W. Fan, P. Qiao, Vibration-based damage identification methods: a review and comparative study. Struct. Health Monit. 10(1), 83–111 (2011)Google Scholar
  22. 22.
    S.W. Doebling, C.R. Farrar, M.B. Prime et al., Damage identification and health monitoring of structural and mechanical systems from changes in their vibration characteristics: a literature review (Los Alamos, Los Alamos National Lab, 1996)Google Scholar
  23. 23.
    E.P. Carden, P. Fanning, Vibration based condition monitoring: a review. Struct. Health Monit. 3(4), 355–377 (2004)Google Scholar
  24. 24.
    B. Grisso, D. M. Peairs, D. J. Inman, Impedance-based health monitoring of composites. in 22nd, Conference on structural dynamics, pp. 26–29 (Dearborn, MI, 2004)Google Scholar
  25. 25.
    B. Huang, H.S. Kim, Frequency response analysis of a delaminated smart composite plate. J. Intell. Mater. Syst. Struct. 26(9), 1091–1102 (2015)Google Scholar
  26. 26.
    B. Huang, B.-H. Koh, H.S. Kim, PCA-based damage classification of delaminated smart composite structures using improved layerwise theory. Comput. Struct. 141, 26–35 (2014)Google Scholar
  27. 27.
    C.J. Keulen, M. Yildiz, A. Suleman, Damage detection of composite plates by Lamb wave ultrasonic tomography with a sparse hexagonal network using damage progression trends. Shock Vib (2014).  https://doi.org/10.1155/2014/949671 Google Scholar
  28. 28.
    S.M. Prasad, K. Balasubramaniam, C. Krishnamurthy, Structural health monitoring of composite structures using Lamb wave tomography. Smart Mater. Struct. 13(5), N73 (2004)Google Scholar
  29. 29.
    D. Aggelis, N.-M. Barkoula, T. Matikas et al., Acoustic structural health monitoring of composite materials: damage identification and evaluation in cross ply laminates using acoustic emission and ultrasonics. Compos. Sci. Technol. 72(10), 1127–1133 (2012)Google Scholar
  30. 30.
    K. Alnefaie, Finite element modeling of composite plates with internal delamination. Compos. Struct. 90(1), 21–27 (2009)Google Scholar
  31. 31.
    C. Kyriazoglou, B. Le Page, F. Guild, Vibration damping for crack detection in composite laminates. Compos. A Appl. Sci. Manuf. 35(7–8), 945–953 (2004)Google Scholar
  32. 32.
    P. Qiao, K. Lu, W. Lestari et al., Curvature mode shape-based damage detection in composite laminated plates. Compos. Struct. 80(3), 409–428 (2007)Google Scholar
  33. 33.
    C.S. Hamey, W. Lestari, P. Qiao et al., Experimental damage identification of carbon/epoxy composite beams using curvature mode shapes. Struct. Health Monit. 3(4), 333–353 (2004)Google Scholar
  34. 34.
    B. Huang, H.S. Kim, Transient analysis of biocomposite laminates with delamination. J. Nanosci. Nanotechnol. 14(10), 7432–7438 (2014)Google Scholar
  35. 35.
    H.S. Kim, A. Chattopadhyay, A. Ghoshal, Characterization of delamination effect on composite laminates using a new generalized layerwise approach. Comput. Struct. 81(15), 1555–1566 (2003)Google Scholar
  36. 36.
    M. Sause, A. Gribov, A. Unwin et al., Pattern recognition approach to identify natural clusters of acoustic emission signals. Pattern Recogn. Lett. 33(1), 17–23 (2012)Google Scholar
  37. 37.
    T. Ely, E. Hill, Longitudinal splitting and fiber breakage characterization in graphite epoxy using acoustic emission data. NDT E Int. 2(30), 109 (1997)Google Scholar
  38. 38.
    M. Suzuki, H. Nakanishi, M. Iwamoto et al., Application of static fracture mechanisms to fatigue fracture behavior of class A-SMC composite. in Japan-US conference on composite materials, 4 th, Washington, DC (1989), pp. 297–306Google Scholar
  39. 39.
    R. Gutkin, C. Green, S. Vangrattanachai et al., On acoustic emission for failure investigation in CFRP: pattern recognition and peak frequency analyses. Mech. Syst. Signal Process. 25(4), 1393–1407 (2011)Google Scholar
  40. 40.
    J.J. Valletta, C. Torney, M. Kings et al., Applications of machine learning in animal behaviour studies. Anim. Behav. 124, 203–220 (2017)Google Scholar
  41. 41.
    D. Kateris, D. Moshou, X.-E. Pantazi et al., A machine learning approach for the condition monitoring of rotating machinery. J. Mech. Sci. Technol. 28(1), 61–71 (2014)Google Scholar
  42. 42.
    S. Park, J.-J. Lee, C.-B. Yun et al., Abuilt-in active sensing system-based structural health monitoring technique using statistical pattern recognition. J. Mech. Sci. Technol. 21(6), 896–902 (2007)Google Scholar
  43. 43.
    H. Wang, C. Ma, and L. Zhou, A brief review of machine learning and its application. in 2009 international conference on information engineering and computer science (2009), pp. 1–4Google Scholar
  44. 44.
    V. Cherkassky, F. Mulier, Learning from data: concepts, theory, and methods, Adaptive and Learning Systems for Signal Processing, Communications and Control Series (Wiley, New York, 1998)zbMATHGoogle Scholar
  45. 45.
    K. Worden, C.R. Farrar, G. Manson et al., The fundamental axioms of structural health monitoring. Proc. Eng. R. Soc. A Math. Phys. Eng. 463, 1639–1664 (2007)Google Scholar
  46. 46.
    J. Ghaboussi, J. Garrett Jr., X. Wu, Knowledge-based modeling of material behavior with neural networks. J. Eng. Mech. 117(1), 132–153 (1991)Google Scholar
  47. 47.
    H. Su, K.T. Chong, Neural network based expert system for induction motor faults detection. J. Mech. Sci. Technol. 20(7), 929 (2006)Google Scholar
  48. 48.
    Y.-R. Hwang, K.-K. Jen, Y.-T. Shen, Application of cepstrum and neural network to bearing fault detection. J. Mech. Sci. Technol. 23(10), 2730 (2009)Google Scholar
  49. 49.
    H. Li, Y. Zhang, H. Zheng, Gear fault detection and diagnosis under speed-up condition based on order cepstrum and radial basis function neural network. J. Mech. Sci. Technol. 23(10), 2780–2789 (2009)Google Scholar
  50. 50.
    K.T. Chong, A.G. Parlos, Comparison of traditional and neural network approaches to stochastic nonlinear system identification. KSME Int. J. 11(3), 267–278 (1997)Google Scholar
  51. 51.
    S.-Y. Kim, B.-Y. Moon, D.-E. Kim, Optimum design of ship design system using neural network method in initial design of hull plate. KSME Int. J. 18(11), 1923–1931 (2004)Google Scholar
  52. 52.
    A.S. Islam, K.C. Craig, Damage detection in composite structures using piezoelectric materials (and neural net). Smart Mater. Struct. 3(3), 318 (1994)Google Scholar
  53. 53.
    A.C. Okafor, K. Chandrashekhara, Y. Jiang, Delamination prediction in composite beams with built-in piezoelectric devices using modal analysis and neural network. Smart Mater. Struct. 5(3), 338 (1996)Google Scholar
  54. 54.
    D.-U. Sung, J.-H. Oh, C.-G. Kim et al., Impact monitoring of smart composite laminates using neural network and wavelet analysis. J. Intell. Mater. Syst. Struct. 11(3), 180–190 (2000)Google Scholar
  55. 55.
    D. Sammons, W. P. Winfree, E. Burke et al., Segmenting delaminations in carbon fiber reinforced polymer composite CT using convolutional neural networks. in AIP conference proceedings, vol. 1706 (AIP Publishing, Melville, 2016) p. 110014Google Scholar
  56. 56.
    H. Bar, M. Bhat, C. Murthy, Identification of failure modes in GFRP using PVDF sensors: ANN approach. Compos. Struct. 65(2), 231–237 (2004)Google Scholar
  57. 57.
    C. Bhat, M. Bhat, C. Murthy, Acoustic emission characterization of failure modes in composites with ANN. Compos. Struct. 61(3), 213–220 (2003)Google Scholar
  58. 58.
    D. Chetwynd, F. Mustapha, K. Worden et al., Damage localisation in a stiffened composite panel. Strain 44(4), 298–307 (2008)Google Scholar
  59. 59.
    C. Cortes, V. Vapnik, Support-vector Networks. Mach. Learn. 20(3), 273–297 (1995)zbMATHGoogle Scholar
  60. 60.
    L. Wang, Support vector machines: theory and applications (Springer Science & Business Media, Berlin, 2005)zbMATHGoogle Scholar
  61. 61.
    H.-Z. Huang, H.-K. Wang, Y.-F. Li et al., Support vector machine based estimation of remaining useful life: current research status and future trends. J. Mech. Sci. Technol. 29(1), 151–163 (2015)Google Scholar
  62. 62.
    G.-M. Lim, D.-M. Bae, J.-H. Kim, Fault diagnosis of rotating machine by thermography method on support vector machine. J. Mech. Sci. Technol. 28(8), 2947–2952 (2014)Google Scholar
  63. 63.
    Z. Xing, J. Qu, Y. Chai et al., Gear fault diagnosis under variable conditions with intrinsic time-scale decomposition-singular value decomposition and support vector machine. J. Mech. Sci. Technol. 31(2), 545–553 (2017)Google Scholar
  64. 64.
    K.-R. Kim, Y.-S. Lee, Acoustic emission source localization in plate-like structures using least-squares support vector machines with delta t feature. J. Mech. Sci. Technol. 28(8), 3013–3020 (2014)Google Scholar
  65. 65.
    Y. Kim, J. Jang, W. Kim et al., Multiple defect diagnostics of gas turbine engine using SVM and RCGA-based ANN algorithms. J. Mech. Sci. Technol. 26(5), 1623–1632 (2012)Google Scholar
  66. 66.
    D.-H. Seo, T.-S. Roh, D.-W. Choi, Defect diagnostics of gas turbine engine using hybrid SVM-ANN with module system in off-design condition. J. Mech. Sci. Technol. 23(3), 677–685 (2009)Google Scholar
  67. 67.
    S. Das, A. Chattopadhyay, A.N. Srivastava, Classifying induced damage in composite plates using one-class support vector machines. Aiaa J. 48(4), 705–718 (2010)Google Scholar
  68. 68.
    P.M. Pawar, S.N. Jung, Support vector machine based online composite helicopter rotor blade damage detection system. J. Intell. Mater. Syst. Struct. 19(10), 1217–1228 (2008)Google Scholar
  69. 69.
    M. Farooq, H. Zheng, A. Nagabhushana et al., Damage detection and identification in smart structures using SVM and ANN. in Smart Sensor Phenomena, Technology, Networks, and Systems Integration 2012, vol. 8346 (International Society for Optics and Photonics, Bellingham, 2012), p. 83461OGoogle Scholar
  70. 70.
    G. Dib, O. Karpenko, E. Koricho et al., Ensembles of novelty detection classifiers for structural health monitoring using guided waves. Smart Mater. Struct. 27(1), 015003 (2017)Google Scholar
  71. 71.
    F. Cong, J. Chen, G. Dong, Spectral kurtosis based on AR model for fault diagnosis and condition monitoring of rolling bearing. J. Mech. Sci. Technol. 26(2), 301–306 (2012)Google Scholar
  72. 72.
    L.D. Avendaño-Valencia, S.D. Fassois, Natural vibration response based damage detection for an operating wind turbine via random coefficient linear parameter varying AR modelling. J Phys 628, 012073 (2015)Google Scholar
  73. 73.
    Z. Wang, K. Ong, Structural damage detection using autoregressive-model-incorporating multivariate exponentially weighted moving average control chart. Eng. Struct. 31(5), 1265–1275 (2009)Google Scholar
  74. 74.
    E.P. Carden, J.M. Brownjohn, ARMA modelled time-series classification for structural health monitoring of civil infrastructure. Mech. Syst. Sig. Process. 22(2), 295–314 (2008)Google Scholar
  75. 75.
    K.K. Nair, A.S. Kiremidjian, K.H. Law, Time series-based damage detection and localization algorithm with application to the ASCE benchmark structure. J. Sound Vib. 291(1–2), 349–368 (2006)Google Scholar
  76. 76.
    D. Nardi, L. Lampani, M. Pasquali et al., Detection of low-velocity impact-induced delaminations in composite laminates using Auto-Regressive models. Compos. Struct. 151, 108–113 (2016)Google Scholar
  77. 77.
    K. Vamvoudakis-Stefanou, J. Sakellariou, S. Fassois, Random vibration response-only damage detection for a set of composite beams. in Proceedings of the ISMA international conference on noise and vibration engineering (2014), pp. 3839–3853Google Scholar
  78. 78.
    G. Manson, K. Worden, K. Holford et al., Visualisation and dimension reduction of acoustic emission data for damage detection. J. Intell. Mater. Syst. Struct. 12(8), 529–536 (2001)Google Scholar
  79. 79.
    L.E. Mujica, J. Vehí, M. Ruiz et al., Multivariate statistics process control for dimensionality reduction in structural assessment. Mech. Syst. Signal Process. 22(1), 155–171 (2008)Google Scholar
  80. 80.
    G. Kerschen, P. De Boe, J.-C. Golinval et al., Sensor validation using principal component analysis. Smart Mater. Struct. 14(1), 36 (2004)Google Scholar
  81. 81.
    G. Manson, “Identifying damage sensitive, environment insensitive features for damage detection. in Proceedings of the third international conference on identification in engineering systems (2002), pp. 187–197Google Scholar
  82. 82.
    A.-M. Yan, G. Kerschen, P. De Boe et al., Structural damage diagnosis under varying environmental conditions—part I: a linear analysis. Mech. Syst. Signal Process. 19(4), 847–864 (2005)Google Scholar
  83. 83.
    A.R. Oskouei, H. Heidary, M. Ahmadi et al., Unsupervised acoustic emission data clustering for the analysis of damage mechanisms in glass/polyester composites. Mater. Des. 37, 416–422 (2012)Google Scholar
  84. 84.
    J.O. Berger, Statistical decision theory and Bayesian analysis (Springer Science & Business Media, New York, 2013)Google Scholar
  85. 85.
    J.M. Bernardo, A.F. Smith, Bayesian theory. 1994 (Wiley, Valencia (España), 1994)Google Scholar
  86. 86.
    J. Yu, M. Bai, G. Wang et al., Fault diagnosis of planetary gearbox with incomplete information using assignment reduction and flexible naive Bayesian classifier. J. Mech. Sci. Technol. 32(1), 37–47 (2018)Google Scholar
  87. 87.
    T. Peng, A. Saxena, K. Goebel et al., A novel Bayesian imaging method for probabilistic delamination detection of composite materials. Smart Mater. Struct. 22(12), 125019 (2013)Google Scholar
  88. 88.
    M. Nguyen, X. Wang, Z. Su et al., Damage identification for composite structures with a Bayesian network. in Proceedings of the 2004 intelligent sensors, sensor networks and information processing conference, 2004 (2004), pp. 307–311Google Scholar
  89. 89.
    E. Rabiei, E. L. Droguett, M. Modarres, Damage monitoring and prognostics in composites via dynamic Bayesian networks. in 2017 Annual reliability and maintainability symposium (RAMS) (2017), pp. 1–7Google Scholar
  90. 90.
    O. Addin, S. Sapuan, E. Mahdi et al., A Naïve-Bayes classifier for damage detection in engineering materials. Mater. Des. 28(8), 2379–2386 (2007)Google Scholar
  91. 91.
    J. Nichols, W. Link, K. Murphy et al., A Bayesian approach to identifying structural nonlinearity using free-decay response: application to damage detection in composites. J. Sound Vib. 329(15), 2995–3007 (2010)Google Scholar
  92. 92.
    N. Godin, S. Huguet, R. Gaertner et al., Clustering of acoustic emission signals collected during tensile tests on unidirectional glass/polyester composite using supervised and unsupervised classifiers. NDT E Int. 37(4), 253–264 (2004)Google Scholar
  93. 93.
    F. Pashmforoush, R. Khamedi, M. Fotouhi et al., Damage classification of sandwich composites using acoustic emission technique and k-means genetic algorithm. J. Nondestr. Eval. 33(4), 481–492 (2014)Google Scholar
  94. 94.
    T.G. Dietterich, Ensemble methods in machine learning (Springer, Berlin, 2000), pp. 1–15Google Scholar
  95. 95.
    C. Zhang, Y. Ma, Ensemble machine learning: methods and applications (Springer, Berlin, 2012)zbMATHGoogle Scholar
  96. 96.
    P. Gaudenzi, D. Nardi, I. Chiapetta et al., Impact damage detection in composite laminate plates using an integrated piezoelectric sensor and actuator couple combined with wavelet based features extraction approach. in Proceedings of the 7th ECCOMAS thematic conference on smart structures and materials, Azores, Portugal (2015)Google Scholar
  97. 97.
    J.P. McCrory, S.K. Al-Jumaili, D. Crivelli et al., Damage classification in carbon fibre composites using acoustic emission: a comparison of three techniques. Compos. B Eng. 68, 424–430 (2015)Google Scholar
  98. 98.
    S. Kessler, P. Rani, Pattern recognition for damage characterization in composite materials. in 48th AIAA/ASME/ASCE/AHS/ASC structures, structural dynamics, and materials conference (2007), p. 2411Google Scholar
  99. 99.
    R. De Oliveira, A. Marques, Health monitoring of FRP using acoustic emission and artificial neural networks. Comput. Struct. 86(3–5), 367–373 (2008)Google Scholar
  100. 100.
    L. Li, S.V. Lomov, X. Yan et al., Cluster analysis of acoustic emission signals for 2D and 3D woven glass/epoxy composites. Compos. Struct. 116, 286–299 (2014)Google Scholar
  101. 101.
    N. Godin, S. Huguet, R. Gaertner, Integration of the Kohonen’s self-organising map and k-means algorithm for the segmentation of the AE data collected during tensile tests on cross-ply composites. NDT E Int. 38(4), 299–309 (2005)Google Scholar
  102. 102.
    F. Pashmforoush, M. Fotouhi, M. Ahmadi, Acoustic emission-based damage classification of glass/polyester composites using harmony search k-means algorithm. J. Reinf. Plast. Compos. 31(10), 671–680 (2012)Google Scholar
  103. 103.
    S.K. Al-Jumaili, K.M. Holford, M.J. Eaton et al., Classification of acoustic emission data from buckling test of carbon fibre panel using unsupervised clustering techniques. Struct. Health Monit. 14(3), 241–251 (2015)Google Scholar
  104. 104.
    D.-A. Tibaduiza, M.-A. Torres-Arredondo, L. Mujica et al., A study of two unsupervised data driven statistical methodologies for detecting and classifying damages in structural health monitoring. Mech. Syst. Signal Process. 41(1–2), 467–484 (2013)Google Scholar
  105. 105.
    J. Vitola, F. Pozo, D.A. Tibaduiza et al., Distributed piezoelectric sensor system for damage identification in structures subjected to temperature changes. Sensors 17(6), 1252 (2017)Google Scholar
  106. 106.
    A.A. Jiménez, C.Q.G. Muñoz, F.P.G. Márquez, Machine learning for wind turbine blades maintenance management. Energies 11(1), 1–16 (2017)Google Scholar
  107. 107.
    J. Tang, G. Yan, Monitoring and detection of structural damage under complex environment using guided waves. in Singapore international NDT conference and exhibition (2013)Google Scholar
  108. 108.
    H. Sohn, G. Park, J.R. Wait et al., Wavelet-based active sensing for delamination detection in composite structures. Smart Mater. Struct. 13(1), 153 (2003)Google Scholar
  109. 109.
    M. Sultan, K. Worden, S. Pierce et al., On impact damage detection and quantification for CFRP laminates using structural response data only. Mech. Syst. Signal Process. 25(8), 3135–3152 (2011)Google Scholar
  110. 110.
    M. Fotouhi, H. Heidary, M. Ahmadi et al., Characterization of composite materials damage under quasi-static three-point bending test using wavelet and fuzzy C-means clustering. J. Compos. Mater. 46(15), 1795–1808 (2012)Google Scholar

Copyright information

© The Korean Society of Mechanical Engineers 2019

Authors and Affiliations

  • Asif Khan
    • 1
  • Nayeon Kim
    • 1
  • Jae Kyong Shin
    • 1
  • Heung Soo Kim
    • 1
    Email author
  • Byeng Dong Youn
    • 2
  1. 1.Department of Mechanical, Robotics and Energy EngineeringDongguk University-SeoulSeoulRepublic of Korea
  2. 2.System Health and Risk Management Laboratory, Department of Mechanical and Aerospace EngineeringSeoul National UniversitySeoulSouth Korea

Personalised recommendations