Elsevier

Composite Structures

Volume 116, September–October 2014, Pages 286-299
Composite Structures

Cluster analysis of acoustic emission signals for 2D and 3D woven glass/epoxy composites

https://doi.org/10.1016/j.compstruct.2014.05.023Get rights and content

Abstract

The goal of this work is to identify the best strategy for clustering of AE events, originated from damage initiation and development of 2D and 3D glass/epoxy woven composites loaded in tension. Two AE features – peak amplitude and peak frequency – were selected as the best cluster-definition features from nine AE parameters by (a) Laplacian score and correlation analysis, (b) principal component analysis and k-means++ algorithm and (c) repeatability and similarity analysis of the clusters in AE registration of different specimens. Peak amplitude and peak frequency represent adequately and in a reproducible way the AE events clustering for both 2D and 3D woven glass/epoxy composites, resulting in the clusters of similar shape. Cluster bounds are identified for different reinforcement type and different loading directions. The cluster identification creates a framework for analysis of a link between damage mode and AE parameters of the corresponding AE event.

Introduction

Acoustic emission (AE) registration allows detection of micro-damage events in materials and structures and potentially identification of the damage nature. An acoustic emission signal is an ultrasonic wave resulting from the sudden release of the strain energy when damage happens, for example after fiber breakage, interface debonding, matrix cracking or delamination in composite materials. AE signals contain useful information about the damage mechanisms. The challenge of AE analysis is determining connection between the AE signal parameters and the corresponding damage mode, or discriminating AE signals according to the nature of the damage events they originate from one of the generally accepted ways of discriminating AE signals is cluster analysis, which is a synonym for unsupervised pattern recognition technique [1], [2], [3], [4], [5], [6], [7], [8].

In cluster analysis the three important things are [4], [5], [6], [9]: (1) selection of AE features to be used for cluster definition, (2) choosing the clustering algorithm and (3) validation of the defined clusters.

For the choice of AE features, many descriptors provide rich information about AE signal characteristics, but non-selective use of them may lead to redundancy information for pattern recognition if strong correlation exists among descriptors. Sause [5] made use of an exhaustive search method in feature selection procedure and considered wide range of combinations of signal features extracted from AE signals. For cluster algorithm, the most frequently used methods are k-means, self-organized map combination with k-means and fuzzy-c means algorithm. K-means algorithm is the simplest and effective method for AE signal clustering. Huguet [10], Godin [11], Oliveira [12] and Gutkin [4] used self-organized map combined with k-means to cluster the AE signals.

For investigation of the cluster validity, the most frequently used indexes are Silhouette coefficient [4] and Davies–Bouldin index [9], [10], [11]. Sause [5] utilized validation of obtained partitions by cluster validity indices (Davies–Bouldin, Tou indices, Rousseeuw’s silhouette validation method and Hubert’s Gamma statistics) and voting scheme for the number of clusters with best performance.

Each AE event can be considered as an acoustic signature of one of different damage modes [10]. AE cluster analysis was applied to identification of damage mechanisms in unidirectional glass fiber reinforced polymer (GFRP) in a number of studies [9], [10], [11], [13], [14], [15], [16]. Some authors used time domain features (peak amplitude, duration, rise time, counts, counts to peak and energy) and found that damage mechanisms are highly correlated with peak amplitude. Barre and Benzeggagh [17] give the following peak amplitude ranges for glass/polypropylene composites: 40–55 dB is matrix cracking, 60–65 dB relates to debonding, 65–85 dB relates to fiber pull-out and 85–95 dB is fiber fracture. Peak amplitude related clusters are also described in [9], [14], [17]. Some researchers [9], [10], [11] utilized multiple AE time features for clustering and concluded that there are four typical types of AE signals for GFRP composites. Fig. 1 illustrates three types of them as seen in the experiments reported here for glass/epoxy composites, with interpretation according to [11]: A-type signal associated with matrix cracking with low amplitude, medium rise time and medium duration, B-type signal related to interface debonding and had short rise time and short duration, D-type signal with medium amplitude, medium rise time and long duration corresponded to delamination in glass/epoxy composites. C-type signals, reported in glass/polyester composites [10] and analogous to B-type, but with much lower amplitude, were not observed in the present study. Ely and Hill [18] showed that in unidirectional graphite/epoxy specimen the stronger signals (high amplitude, energy, counts and long duration) resulted from fiber breakage and the weaker ones (low amplitude, energy, counts and short duration) resulted from longitudinal split cracks. However, this is not supported by the similar observations for glass/epoxy composites.

Some authors used time and frequency features and found that frequency ranges of the clusters are well distinguished, whereas there is an obvious overlap for other parameters [15], [16], [19]. This leads to a hypothesis that a frequency range is representative of a specific damage mechanism, and frequency can be regarded as the best AE descriptor for damage characterization. Table 1 summaries damage mode links to AE features proposed by different authors for glass fiber reinforced composites.

The materials studied in the present paper are glass/epoxy composites, reinforced with four plies of plain weave E-glass fabric (2D) and single-ply non-crimp 3D orthogonal (3D) E-glass woven fabrics (Fig. 2 and some properties of the preforms in Table 2 [20]). Quasi-static tensile mechanical behavior of 3D woven glass/epoxy composites and their 2D plain weave counterparts, used in the present paper was studied in detail in [21], [22], while tensile–tensile fatigue performance was investigated in [20]. For plain weave composites, damage mechanisms include matrix cracking (for the warp direction loading: transverse cracks within fill yarns, longitudinal cracks in the warp direction) and delamination (see Fig. 3(a)). Damage evolution of 3D composite is different because the warp and fill layers are interlaced by z-yarns through thickness. The presence of z-yarns affects the position and the type of induced damage. For 3D textile composites loaded in the warp direction, there are bundle-boundary cracks on z-yarns, transverse cracks in fill direction, cracks in z-yarns, cracks in warp yarns (seen Fig. 3(b)). A similar damage pattern was observed for the fill direction loading. For both 2D and 3D glass woven composites, the following damage types can be identified on the intra-ply scale level: (1) matrix intra-yarn cracks, including transverse cracks on yarn boundary, transverse cracks inside yarns, debonding on the yarn surface parallel to the loading; (2) matrix inter-yarn cracks including transverse and longitudinal matrix cracks in matrix pockets, and (3) fiber failure. On the intra-ply level, delamination appear on later stages of the loading.

AE was recorded during these tests, but was analyzed in [21], [22] using only the cumulative energy of the AE events. In the present work the clustering analysis is performed on the AE data registered in an independent series of tests performed during the fatigue study of the same materials [20], [23]. The synopsis of the paper content is as follows: four AE features were selected after the correlation analysis and Laplacian score criteria from the primary nine AE features, and further narrowed to two: peak amplitude and peak frequency of the signal for the cluster analysis. For all specimens the repeatability of AE events peak amplitude and peak frequency shows that the selected features are uniform for all tensile tests. Then k-means++ clustering algorithm and the principal component analysis were used for cluster the AE events for all specimens. The cluster bonds are identified for the two materials (2D and 3D woven) and different loading directions (warp and fill), and generalized AE cluster structure is presented and possible relation of the clusters to the damage types is hypothesized. These results create a foundation for establishment of general AE interpretation rules for damage mode identification in the future work.

Section snippets

Cluster analysis methods

Feature selection is a procedure of extracting the features which are good for classification. ‘Good features’ are such that objects from the same class have similar feature values and objects from different classes have different values. The goal of feature selection is to find the subset of parameters which eliminate irrelevant and redundant features while keeping relevant features in order to improve clustering efficiency and quality. Existence of irrelevant features in the data set may

Typical AE registrations and AE signal parameters

In this study acoustic emission data of 2D and 3D glass/epoxy woven composites recorded in the previously reported experiments [23], [20] is used. Vallen-5 AE acquisition software was used, with the parameters shown in Table 3. Papers [21], [22] report a comparative study of the mechanical properties and damage observation for the two type of materials here considered: four-ply plain weave laminates and single-ply 3D glass non-crimp orthogonal woven composite, which were prepared in the same

Discussion

In order to primarily correlate the resulted clusters with different damage mechanisms, four clusters were compared with peak amplitude distribution and peak frequency band which represents different damage mechanisms in glass fiber composite materials in literature [9], [14], [13], [31]. CL1 and CL2 with amplitude range 35–55 dB have similar distribution with A-type signal as discussed in [11], [14], which corresponds to matrix cracking. Similarly, CL3 with amplitude distribution 55–100 dB

Conclusion

The cluster analysis of AE during tensile loading of 2D and 3D orthogonally woven E-glass/epoxy composites, loaded in the direction of warp or weft fibers leads to the following conclusions valid for both types of the reinforcements and both warp and weft loading:

  • 1.

    AE events can be discriminated in four clusters based on peak amplitude, peak frequency, frequency of centroid and RA value. Peak amplitude (PA) and peak frequency (PF) are the most important parameters in this discrimination.

Acknowledgements

The research visit of L. Li to KU Leuven was funded by Chinese Scholarship Council and partially supported by FWO project G.0354.09. The raw AE data used in the present study was acquired by Guilia Gramellini during her Master thesis [23] research (supervisors V. Carvelli and S.V. Lomov). The help with Vallen AE system of Johan Vanhulst and useful discussions with Dr Helge Pfeiffer are acknowledged with gratitude.

References (31)

Cited by (0)

View full text