Defect Characterisation of GFRP Cross Ply Laminates using Artificial Neural Networks

Defect Characterisation of GFRP Cross Ply Laminates using Artificial Neural Networks

1. K. Karthik, S. Sivaraman

   Assistant Professor, Department of Aeronautical Engineering,

   Hindusthan College of Engg & Technology, Coimbatore.

   P. Balaguru

   Assistant Professor, Department of Aeronautical Engineering,

   Raja Lakshmi Engineering College, Chennai

   R. Prabu

   Assistant Professor, Department of Aeronautical Engineering,

   KCG College of Technology, Chennai

   Abstract An experimental work has been carried out to characterise the defects of post impacted Glass/Epoxy composite laminates using online acoustic emission (AE) monitoring and artificial neural networks (ANN). The laminates were made from ten-layered glass fibre (200 MIL cloth) with epoxy as the binding medium by hand lay-up technique and cured at a pressure of 100 kg/cm2 under room temperature using a 30 ton capacity compression moulding machine for 24 hours. 25 test specimens (ASTM D3039 standard) were prepared from the cross-ply laminates using water jet cutting machine. 21 specimens were subjected to impact load from three different heights using CEAST Fractovis Drop Impact machine. Both impacted and non-impacted specimens were subjected to uniaxial tension under the acoustic emission monitoring using 30 kN INSTRON 3367 universal testing machine. The dominant AE parameters such as counts, energy, duration, rise time and amplitude are recorded during monitoring. These AE parameters are then used to characterise the defects in composite materials using Fuzzy C-means clustering algorithm associated with Principal Component Analysis. Artificial Neural Network technique is used in the process of getting the results. The acquired results can be used for online health monitoring through which failure of composite components can be identified at the initial stages.

   Keywords Acoustic emission monitoring, artificial neural network, Fuzzy C means clustering, principal component analysis, online health monitoring, Glass/Epoxy, Cross-Ply laminates.

   1. INTRODUCTION

      Glass Fiber Reinforced Polymer (GFRP) laminates are widely for aerospace applications because of their high strength to weight ratio and corrosion resistance. However one of the major limitations is the effect of impact damage which leads to multiple level delamination through the thickness. To prevent the material from failing structurally we have to identify the defects at earlier stages itself.

      Acoustic Emission (AE) [4] can be used a tool for the evaluation of damages in fiber reinforced composites. AE cannot be considered as a non-destructive testing method strictly, because the changes which occur during AE are mostly irreversible. In conventional AE investigation, various AE parameters are analyzed directly as a plot of two or three AE parameters, which is not sufficient for composite materials.

      Artificial Neural Network (ANN) [1] analogous to biological neural system is an adaptive computer program which provides solutions to problems like complex data collections. In this approach relationship between input and output parameters are developed through a training process in which sets of inputs are applied to the network and the resulting outputs are compared with the known results. ANNs are trained using either supervised or unsupervised learning models. The trained networks are used to predict the output that would result from a set of inputs which are not included in the training data.

      The major challenge with data analysis is the discrimination between the different acoustic emission sources. The objective of the cluster analysis is to separate a set of data into several classes that reflect the internal structure of the data. Indeed, cluster analysis is an important tool for investigating and interpreting data.

      In order to improve the cluster analysis process, fuzzy c-means clustering associated with a principal component analysis are proposed in this paper. The fuzzy c-means clustering method (FCM) is an effective unsupervised algorithm for automatic clustering and separating AE patterns composed of multiple features extracted from the random AE waveforms. The five descriptors used are the energy, amplitude, rise time, counts and duration of the AE signals. FCM is applied to the above data and the damage mechanisms[7] such as matrix cracking, fiber matrix debonding and delamination signals.

   2. PREPARATION OF TEST SPECIMENS

      Uni-directional glass fibers (S-Class) of dimension 300×300 mm are cut from the big roll.10 such glass fibers are required for preparing a GFRP laminate. The weight of all the

      10 glass fibers is measured using an electronic weighing machine. Epoxy resin equal in weight to that of fiber is weighed and taken separately. The hardener is added to the resin in the ratio of 1:10. The epoxy resin mixture is then mixed thoroughly.

      Place the mould on the table. Apply a thin layer of resin on the surface of the lower mould. Next place the first layer of glass and use rollers to squeeze the excess resin. Apply resin over the first layer of glass and place then place the second glass layer and again use the rollers to squeeze the excess resin.

      Repeat the procedure with alternating layers of glass fiber and resin mixture until all the 10 layers of glass fibers are finished. Place the upper mould over the lower mould and the mould is closed. The mould is placed in the compression moulding machine and cured at a pressure of 100 bar for 24 hours.

      Fig.1 GFRP Cross ply laminate after curing from Compression Moulding Machine

      The GFRP cross ply laminate will be having a thickness of about 1.8mm. 25 test specimens were cut from the laminate using water jet cutting machine according to ASTM D3039 standard.

      Fig.2 ASTM D3039 Standard

   3. INTRODUCTION OF IMPACT DAMAGE IN

      SPECIMEN

      The specimens prepared from the laminates are grouped into three groups of seven specimens each. Each group of specimens was subjected to impact[5],[6] using CEAST Fractovis drop impact machine from three different heights of 75mm, 100mm and 125mm respectively. The parameters

      used during the impact process are: Clamping force 1000 N, Impactor Striker mass 1.926 kg, Impactor diameter 12.7mm. The remaining specimens are kept without impact damage.

      Fig.3 Impact Energy Vs. time

      Fig.4 Post impacted specimen

   4. EXPERIMENTAL PROCEDURE FOR TESTING The impacted specimens are then subjected to uni-axial

      tension using 30kN INSTRON 3367 Universal Testing Machine under acoustic emission monitoring. The parameters used for the tensile loading process are: cross head speed 0.1mm/min, sampling rate 30pts/min. Aluminum tabs [2] are placed at the ends to provide grip as well for noise reduction during AE monitoring.

   5. DATA ACQUISITION USING AE

      When the structure is loaded transient elastic waves are generated due to the rapid release of strain energy within the materials called as acoustic emission. These acoustic emission has various characteristics such as amplitude, rise time, duration, energy, counts, counts to peak, threshold. These characteristics are recorded and used for analysis purpose. An

      8-channel acoustic emission setup along with SAMOS E3.10 data acquisition system supplied by physical acoustic corporation was used. AE measurement were made using two

      Nano 30 PAC sensors attached to the specimen through high vacuum grease couplant. The signals from the transducer are then passed through PAC 2/4/6 G/A pre-amplifier before reaching the main unit. The input parameters[9] used or AE monitoring are: peak definition time (PDT) 30µs, hit definition time (HDT) – 300µs , hit lock-out time (HLT) – 600µs. AE wave velocity was found as 3149523.8 m/s by Hsu-Nielson source (pencil lead break test).

      Fig.5 Amplitude Vs Location Plot for the specimen impacted from 75mm height

   6. CLUSTERING USING ANN

      1. Multivariable Data Clustering:

         Acoustic Emission Signals are used for the investigation of local damage in composite materials. The problem in the analysis of the AE signals is to identify the most critical damage mechanisms. The unsupervised pattern recognition analysis [8] associated with a principal component analysis is a tool that is used for the classification of the monitored AE events. A cluster analysis of the AE data is achieved and the resulting clusters are correlated to the damage mechanisms of the material under investigation.

         Different damage mechanisms have been identified on fiber matrix composite materials from their AE signals. According to the previous studies, the damage mechanisms that are considered according to the collected AE signals are matrix cracking, fiber cut, and delamination signals. Since there are three damage mechanisms that occur within the composites, the classification to make is considered as a three

         class problem.

         Fuzzy c-means clustering method is thus applied with three clusters. The five descriptors used are the rise time, count, energy, duration, amplitude. Next Principal component analysis is achieved in order to visualize the results in a two dimension subspace.

      2. Principal Component Analysis

         One of the difficulties in multivariate statistics is the problem of visualizing data that has many variables. Principal Component Analysis is a quantitatively rigorous method for achieving this simplification. The method generates a new set of variables, called principal components [3]. Each principal component is a linear combination of the original variables. All the principal components are orthogonal to each other, so there is no redundant information. The principal components as a whole form an orthogonal basis for the space of the data.

         The first principal component is a single axis in space. When you project each observation on that axis, the resulting values form a new variable. And the variance of this variable is the maximum among all possible choice of the first axis. The Second principal component is another axis in space, perpendicular to the first. Projecting the observations on this axis generates another new variable. The variance of this variable is the maximum among all possible choices of this second axis.

         The full set of principal components is as large as the original set of variables. But it is commonplace for the sum of the variances of the first few principal components to exceed 80% of the total variance of the original data.

      3. Fuzzy C-Means Clustering

      Fuzzy C-means clustering algorithm divides out an input data among a pre-defined number k of classes. The classification criterion is then the minimization of the sum of the squared distances between all the descriptor vectors of a cluster and its center. At the beginning, the coordinates of the cluster centres are initialized and each descriptor vector is assigned to the cluster whose center is the closest.

      Fuzzy c-means (FCM) is a data clustering technique wherein each data point belongs to a cluster to some degree that is specified by a membership grade. This technique was originally introduced by Jim Bezdek in 1981 as an improvement on earlier clustering methods. It provides a method that shows how to group data points that populate some multidimensional space into a specific number of different clusters.

      The fuzzy logic starts with an initial guess for the cluster centers, which are intended to mark the mean location of each cluster. The initial guess for these cluster centers is most likely incorrect. Additionally it assigns every data point a membership grade for each cluster. By iteratively updating the cluster centers and membership grades for each point, it iteratively moves the cluster centers to the right location within a data set. The iteration is based on minimizing an objective function that represents the distance from any given data point into a cluster center weighted by that data points membership grade.

      Fig.6 Plot between first and second principal components.

      Fig.7 Plot showing the clusters formed using fuzzy c-means clustering

   7. RESULTS AND DISCUSSION

1. Impact test results:

   Impact test was done using CEAST Fractovis Drop Impact machine at various height and the impact energy for each height is tabulated below,

   IMPACT ENERGY (J)
   Spec. No	Impact Energy (J)	Spec. No	Impact Energy (J)
   75_2	1.346	100_5	1.838
   75_3	1.363	100_8	1.922
   75_4	1.379	100_9	1.795
   75_8	1.460	125_2	2.15
   75_9	1.392	125_5	2.21
   75_10	1.382	125_6	2.25
   75_11	1.375	125_7	2.25
   100_1	1.729	125_8	2.26
   100_2	1.757	125_10	2.39
   100_3	1.826	125_11	2.31
   100_4	1.838

   Table1. Impact Energy Results

2. Tensile test results:

   The impacted specimens are then subjected to uni axial tensile load using INSTRON 3367 Universal Testing Machine and the failure loads of all the specimens are tabulated below,

   FAILURE LOAD (kN)
   Spec. No	Failure Load	Spec. No	Failure Load
   75mm2_2	3.386	100mm5_5	2.976
   75mm3_3	3.692	100mm8_8	3.197
   75mm 4_4	3.847	100mm9_9	2.719
   75mm 8_8	3.426	125mm2_2	2.213
   75mm 9_9	3.591	125mm5_5	1.895
   75mm 10_10	3.562	125mm6_6	1.808
   75mm 11_11	3.779	125mm7_7	1.975
   100mm1_1	3.042	125mm8_8	2.215
   100mm2_2	2.728	125mm10_10	1.728
   100mm3_3	2.899	125mm11_11	2.285
   100mm4_4	3.118

   Table2. Tensile Failure Loads

3. AE Data Acquisition results

   During Tensile test, AE setup was used for data acquisition of various AE parameters such as Amplitude, Duration, Rise time, Energy, Counts etc., The AE data obtained during the test should be then clustered and sorted into three different groups, each group representing the failure mechanism of composite. The following tables represent the total number of events recorded during each test and the clustered data representing the failure mechanisms.

   EVENT DATA HISTORY
   Spec. No	TOTAL EVENTS	CLASS-I	CLASS-II	CLASS-III
   75_2	4084	3342	705	37
   75_3	7256	6582	626	48
   75_4	1958	1720	232	6
   75_8	2690	2676	13	1
   75_9	4194	3375	774	45
   75_10	3338	3018	312	8
   75_11	6932	6803	124	1
   100_1	8128	5627	2414	87
   100_2	11868	11737	123	8
   100_3	3918	2907	974	37
   100_4	3866	3864	1	1
   100_5	7076	6982	92	2
   100_8	2340	2336	3	1
   100_9	9334	7963	1281	103
   125_2	16323	15376	946	1
   125_5	4566	2951	1552	63
   125_6	4658	3510	1098	50
   125_7	12707	12680	24	3
   125_8	7776	7752	23	1
   125_10	3398	2614	698	86
   125_11	3922	3913	8	1

   Table3. AE Date Acquisition

4. Cluster Analysis results:

   Cluster analysis is performed to group the acquired AE data into different clusters. Once the AE data are clustered, then the various AE parameters are analyzed and their ranges are tabulated which are given in the tabular column listed below,

   AE PARAMETER RANGE OF CLASS-I SIGNALS:
   Spec. No	TOTAL EVENT	RISE	COUNT	ENERGY	DURATION	AMP
   75_2	354	1-268	20-107	1-46	74-474	51-79
   75_3	6582	1-149	20-48	2-45	68-197	52-86
   75_4	1720	1-121	20-53	2-39	73-203	50-84
   75_8	2676	1-587	20-179	2-84	61-960	54-85
   75_9	3375	1-99	20-47	2-41	75-175	52-85
   75_10	3018	1-247	20-64	2-51	79-271	54-85
   75_11	6803	1-270	20-193	2-94	65-1210	51-86
   100_1	8128	1-136	20-41	2-47	66-167	54-86
   100_2	11868	1-1293	20-610	2-183	63-3197	51-95
   100_3	3918	1-119	20-40	2-34	69-167	53-84
   100_4	3866	1-642	20-721	2-735	69-3134	53-91
   100_5	7076	1-396	20-74	2-161	68-405	52-93
   100_8	2340	1-775	20-279	2-156	78-1565	52-81
   100_9	9334	1-161	20-47	2-29	69-182	53-80
   125_2	16323	1-145	20-638	2-387	59-3063	53-86
   125_5	4566	1-95	20-38	2-27	61-151	56-82
   125_6	4658	1-108	20-41	2-38	71-175	57-86
   125_7	12707	1-713	20-189	2-313	60-1030	51-95
   125_8	7776	1-651	20-216	2-207	64-1232	52-94
   125_10	3398	1-582	20-176	2-58	63-941	54-83
   125_11	3922	1-558	20-354	2-220	73-1904	54-89

   Table4. Clustering for Class-1 signals

   TYPES OF SIGNAL	AE PARAMETERS
   	RISE	COUNT	ENERGY	DUR	AMP
   CLASS-I	1 120	20 50	2 45	68 203	52 85
   	1 600	20 200	2 200	65 1500	51 90
   CLASS-II	1 200	20 100	3 125	150 450	55 90
   	1 2500	100 1200	50 1000	500 6000	60 100
   CLASS-III	1 1000	55 850	18 500	350 4500	65 90
   	>2500	>6000	>10000	>25000	>99

   1-426

   AE PARAMETER RANGE OF CLASS-II SIGNALS:
   Spec. No	TOTAL EVENT	RISE	COUNT	ENERGY	DURATION	AMP
   75_2	13	11- 1153	99-313	22-134	509-1554	65-82
   75_3	626	1-306	22-143	3-141	176-614	53-95
   75_4	232	25-208	4-197	184-884	53-88
   75_8	13	4-1861	146-611	29-337	973-3099	64-87
   75_9	774	1-242	20-90	3-84	163-436	52-91
   75_10	312	1-723	24-235	4-137	242-1236	55-88
   75_11	124	1-1852	41-626	7-439	344-3199	55-90
   100_1	2414	1-228	20-79	3-96	161-356	53-90
   100_2	123	1-1799	94-611	16-1381	581-2892	60-99
   100_3	974	1-317	20-86	3-209	163-515	53-94
   100_4	1	3914	4175	357	20614	94
   100_5	92	1-1146	46-413	11-386	346-1922	62-98
   100_8	3	1378- 1645	587-1380	232-892	3322-7625	71-90
   100_9	1281	1-224	21-94	3-92	170-453	55-88
   125_2	946	1-3038	21-1172	4-653	182-6260	56-94
   125_5	1552	1-172	20-89	4-231	146-377	58-99
   125_6	1098	1-198	20-91	5-152	167-398	61-93
   125_7	24	44- 8902	147-2261	34-1562	1053-11232	62-99
   125_8	23	125- 3832	196-906	85-857	1370-4757	67-99
   125_10	698	1-159	20-64	3-57	151-334	55-85
   125_11	8	408- 3198	423-1234	194-1156	2080-6003	75-89

   Table7. Defect Characterization Results

   Table5. Clustering for Class-2 signals

   AE PARAMETER RANGE OF CLASS-III SIGNALS:
   Spec. No	TOTAL EVENT	RISE	COUNT	ENERGY	DURATION	AMP
   75_2	1	3534	1399	347	6093	99
   75_3	48	1-2451	74-827	18-445	626-4371	61-90
   75_4	6	7-917	87-309	23-336	836-1587	64-86
   75_8	1	3977	8164	5521	37650	98
   75_9	45	4-1507	36-411	9-249	399-1992	57-89
   75_10	8	2-1700	251-755	55-650	1765-3342	70-88
   75_11	1	875	4402	16668	20003	99
   100_1	87	1-1916	33-753	7-980	321-3759	58-98
   100_2	8	449- 3676	809-2048	306-2430	4068-8331	78-99
   100_3	37	3-695	64-268	16-434	495-1456	60-93
   100_4	1	7676	12534	59731	55511	99
   100_5	2	2742- 2920	3583- 4590	9546-13911	17274-21423	99
   100_8	1	210	4603	12446	25162	99
   100_9	103	1-1300	45-824	12-259	399-4628	63-88
   125_2	1	9538	6849	30456	29851	99
   125_5	63	1-864	55-586	26-513	355-2429	66-99
   125_6	50	1-567	45-379	14-700	361-1775	67-92
   125_7	3	1680- 8259	8513- 14070	3497-25678	38737-59937	82-99
   125_8	1	2576	3646	2220	17309	98
   125_10	86	2-550	28-239	11-113	313-1379	62-86
   125_11	1	9943	7739	30357	34005	99

   Table6. Clustering for Class-3 signals

5. Defect Characterization results:

After performing cluster analysis on the data acquired using Acoustic emission monitoring system during the tensile test performed on the specimen impacted at various energy levels, we obtain three different classes of signals [10] which can be distinguished from each other with the help of AE parameters like, Rise time, Count Energy, Duration and Amplitude. The summary of cluster analysis is given below,

Fig.8 AE Characteristic Signal for Matrix Cracking Failure mode.

Fig.9 AE Characteristic Signal for Delamination Failure mode.

Fig.10 AE Characteristic Signal for Fiber failure mode

7. CONCLUSION AND FUTURE SCOPE:

This experimental work was conducted with the aim of creating a database which contains the AE data for a variety of parameters such as impact load, bending load, fatigue load, compression load, and artificial introduction of defects into laminates so as to understand the initiation and growth pattern of various defects. Once these data are collected for different types of mechanical properties, the data can be kept in a centralized place. Using these centralized data online health monitoring can be performed so as to identify the defects in the initial stages itself and necessary actions can be taken to prevent the failure of structures.

Though this experimental work is limited only with impacted specimens subjected to uni-axial tensile loading, the same process followed in this experiment can be used for all the testing of other mechanical properties and recorded in a centralized place through which only online health monitoring[11] can be achieved.

REFERENCES

1. H.N.Bar, M.R.Bhat, and C.R.L.Murthy Identification of failure modes in GFRP using PVDF sensors: ANN approach in Composite structures 65, 2004, pp. 231 237.

2. C.R. Ramirez-Jimenez, N. Papadakis, N. Reynolds, T.H. Gan, P. Purnell, M. Pharaoh, Identification of failure modes in glass/polypropylene composites by means of the primary frequency content of the acoustic emission event in Composites Science and Technology, 64, 2004, pp.18191827.

3. N. Godina, S. Hugueta, R. Gaertnera, L. Salmon, Clustering of acoustic emission signals collected during tensile tests on unidirectional glass/polyester composite using supervised and unsupervised classifiers in NDT&E International, 37, 2004, pp. 253 264.

4. S. Hugueta, N. Godina, R. Gaertnera, L. Salmon, D. Villard Use of acoustic emission to identify damage modes in glass fiber reinforced polyester in Composites Science and Technology, 62, 2002, pp. 1433

   1444.

5. Steve.E.Watkins, Farhad Akhavan, Rohit dua, Donald C Wunch, Impact Induced Damage Characterization Of Composite Plates using Neural Networks in Smart Mater. Struct., 16, 2007, pp. 515 524.

6. Carlo Santulli, Post-impact damage characterisation on natural fibre reinforced composites using acoustic emission in NDT&E International, 34, 2001, pp. 531 536.

7. Bradford H Parker, Acoustic emission monitoring of low velocity impact damage in graphite/epoxy laminates during tensile loading in NASA Technical Memorandum 4339, 1992.

8. N.Godin, S.Huguet, R.Gaertner, Integration of kohonen's self- organising map and k-means algorithm for the segmentation of the AE data collected during tensile tests on cross-ply laminates in NDT&E International, 38, 2005, pp. 299 309.

9. SAMOS AE System Users Manual, Rev 2, Part#: 7030 1001, 2005

10. R. de Oliveira, A.T. Marques, Damage Mechanisms Identification in FRP using Acoustic Emission and Artificial Neural Networks in Material Science Forum, 514-516, 2006, pp. 789 793.

11. Victor giurgiutiu, Active sensors for health monitoring of aging aerospace structures in SPIEs 7th International Symposium on smart structures and materias and 5th International Symposium on non destructive evaluation and health monitoring of aging infrastructure, 2000, paper # 3985-103.