Automatic ultrasonic inspection for internal defect detection in composite materials

Transcription

1 ARTICLE IN PRESS NDT&E International 41 (2008) Automatic ultrasonic inspection for internal defect detection in composite materials T. D Orazio a,, M. Leo a, A. Distante a, C. Guaragnella b, V. Pianese c, G. Cavaccini c a Istituto di Studi sui Sistemi Intelligenti per l Automazione CNR, via Amendola 122 D/I, Bari, Italy b DEE Politecnico di Bari, Italy c Alenia Aeronautica Pomigliano, Napoli, Italy Received 19 December 2006; received in revised form 16 July 2007; accepted 2 August 2007 Available online 14 August 2007 Abstract The detection of internal defects in composite materials with non-destructive techniques is an important requirement both for quality checks during the production phase and in-service inspection during maintenance operations. Visual inspection allows only the analysis of surface characteristics of materials and, then, if internal faults occur inside composite structures, a deeper analysis is required. A comparison between the reactions of different materials to ultrasonic signals can be used to highlight the difference in the internal structures and also to detect the depth position of these anomalies. However, ultrasonic data are difficult to interpret since they require the analysis of a continuous signal for each point of the material under consideration. An automatic procedure is necessary to manage large data sets and to extract significant differences between them. In this paper, we address the problem of automatic inspection of composite materials using an ultrasonic technique. We consider two main steps for interpreting ultrasonic data: the pre-processing technique necessary to normalize the signals of composite structures with different thicknesses and the classification techniques used to compare the ultrasonic signals and detect classes of similar points. r 2007 Elsevier Ltd. All rights reserved. Keywords: Composite materials; Ultrasonic inspection; Neural network; Signal normalization 1. Introduction The problem of detecting internal defects in composite materials has received great attention in recent years both for quality control during production phases and for inservice inspection during maintenance operations. The use of non-destructive techniques is necessary for the analysis of internal properties of structures without causing damage to the materials. Some of these NDT&E techniques are based on analysis of the transmission of different signals such as ultrasonics, acoustic emission, thermography, laser ultrasonics, X-radiography, eddy currents, shearography, and low frequency methods [1]. In the last decade, ultrasonic techniques have shown to be very promising for non-destructive inspection and they are becoming an effective alternative to such Corresponding author. address: dorazio@ba.issia.cnr.it (T. D Orazio). traditional and well-studied approaches as thermography, eddy current, and shearography. In Ref. [2], an electro-magnetic acoustic transducer (EMAT) working at a small standoff was applied to detection and gauging in real time of surface defects in rail tracks allowing the lifting of many of the previous rail testing speed limitations with a high degree of accuracy. The authors pointed out the relationship between defects and both time and frequency domains of acoustic signals. In Ref. [3], a high resolution pursuit (HRP) based signal processing method was proposed for detecting flaws close to the surface of strongly scattering materials, such as steel and composites, in NDT applications. In Ref. [4], an approach to non-destructive pipeline testing using ultrasonic imaging was proposed. The authors trained a generalized regression neural network to determine the dimensions of the corrosion and to generate the whole image of both the internal and external walls of the oil pipeline. As an improvement to the detection algorithm, /$ - see front matter r 2007 Elsevier Ltd. All rights reserved. doi: /j.ndteint

2 146 ARTICLE IN PRESS T. D Orazio et al. / NDT&E International 41 (2008) they introduced fuzzy decision-based neural network algorithms for the detection and classification of the corrosion. The simulation and experimental results showed that these new methods out-perform the existing methods. In Ref. [5], the wavelet transform has been successfully used in experiments to suppress noise and enhance flaw location using ultrasonic signals, with a good defect localization. The obtained result was then fed to an automatic Artificial Neural Network classification and learning algorithm of defects from A-scan data. In Ref. [6], the authors covered the use of different flaw-detection methods comparing them with their proposed one. The experiment aimed to determine whether a given ultrasonic signal contains a flaw echo or not. The proposed method is based on radial basis function networks, one of the most powerful neural network techniques. This signal-processing technique tried to find the optimal decision criterion. This method compares well with thresholding-based ones, showing an improvement in the probability of false alarms of over 25 30%. So their new method is a good alternative for flaw-detection problems. In paper [7], an evaluation of various types and configurations of neural network developed for the purpose of assisting in accurate flaw detection in steel plates is illustrated. The presented research was conducted using a wide range of samples, including non-defective plates, plates with side-drilled holes, different inclusions and porosity, together with smooth and rough cracks. The obtained results indicate that significant benefits may be obtained from the techniques demonstrated. In this paper, we address the problem of developing an automatic system for the analysis of ultrasonic data in order to detect and classify internal defects in composite materials. We consider two main steps for interpreting ultrasonic data: the pre-processing technique necessary to normalize the signals from composite structures with different thicknesses and the classification techniques used to compare ultrasonic signals and to detect classes of similar points. The second step was carried out by using a multilevel neural approach firstly defective areas are separated from the sound ones and then they are classified on the basis of the defect type and localization in depth. The rest of the paper is structured as follows. In Section 2, an overview of the proposed system is presented, in Section 3, the pre-processing step is detailed, in Section 4, the neural network for classification of defects is described and, finally, in Section 5, the experimental results on three different composite materials are reported. 2. System overview Ultrasonic inspection uses sound signals at frequencies beyond human hearing (more than 20 khz), to estimate some properties of the irradiated material by analyzing either the reflected (reflection working modality) or transmitted (transmission working modality) signals. A typical ultrasonic-based inspection system consists of several functional units such as pulser, receiver, transducer, and display devices. A pulser is an electronic device that can produce high voltage electrical pulse. Driven by the pulser, the transducer generates high-frequency ultrasonic energy. The sound energy is introduced into and propagates through the materials in the form of waves. In the transmission modality, the receiver is placed on the opposite side of the material from the pulser whereas in the reflection modality the pulser and the receiver are placed on the same side of the material. Inspection devices can be or not be in contact with the material. In the latter case, a liquid (couplant) is used to facilitate the transmission of ultrasonic vibrations from the transducer to the test surface. Ultrasonic data can be collected and displayed in a number of different formats. The three most common formats are known in the NDT world as A-scan, B-scan, and C-scan presentations. Each presentation mode provides a different way of looking at and evaluating the region of material being inspected. In this paper, the analysis of ultrasonic data acquired from the reflection working modality and A-scan representation is reported. This means that for each point of the composite material, we have a continuous signal that represents the amount of received ultrasonic energy as a function of time. Fig. 1 shows a logical scheme of the proposed approach. Bearing in mind that the shape of the received ultrasonic signals strongly depends on the thickness of the inspected materials, the first step towards the implementation of an automatic inspection system is to introduce a proper normalization in order to minimize this problem. The proposed normalization technique consists of a procedure capable of suppressing non-significant samples from deeper material until the length of the ultrasonic signal becomes equal to that of those signals emitted from thinner samples. A two-level set of neural classifiers was used to process normalized data. The first level consists of a neural network trained to segment defective areas from sound areas (defect-detection phase). This step produces a binary image containing the defect areas on which successive processing has to be carried out. The second level consists of three neural networks trained to determine three different characteristics: the defect position (it could be placed on the Top, in the Middle or on the Bottom of the considered material), the defect type, and both the displacement and the type of defect in a single step. 3. Data pre-processing In order to develop an automatic system for the analysis of ultrasonic signals, we must either be certain that the considered materials have the same thickness or provide a pre-processing phase during which the signals are normalized allowing comparison between different length signals. Thus, the first problem we have solved in this paper was the

3 ARTICLE IN PRESS T. D Orazio et al. / NDT&E International 41 (2008) Fig. 1. A scheme of the proposed approach. normalization of ultrasonic data coming from different thicknesses of material. Different thicknesses of material have different delays between Frontal echo and Back echo. It is therefore necessary to align signals that relate to materials of different thicknesses. The distance between the Back echo and the Frontal echo can be mathematically computed if the setup parameters (distance of the probe from the standard, sample frequency, compliant type, etc.) and the thickness of the material are known. Assuming that this information is available (thickness could be extracted from planning documentation), we can analyze only the portion of signal between Frontal echo and Back echo. The challenge is to modify the two signals in order to have the same length. We have developed a normalization technique which aligns all the signals to the length of that signal which corresponds to the minimum thickness of the analyzed components. The main objective was to eliminate a number of samples in the portion of the signal that was not significant for the defect-type recognition, without altering the shape of the peaks. In a first approach, we used interpolation procedures to align the signals to the longest one. It was evident that the peak shape was altered with respect to the original signal and thus the successive classification procedure for the defect-type recognition was compromised. To overcome this problem, we have developed a normalization procedure to suppress, with a uniform step, the samples which are under a selected threshold value. The procedure slides the signal from the left to the right and extracts the positions of points having values under the threshold, then evaluates the number of points that have to be eliminated (the difference between the signal length and the desired length of the reference signal). It scans the extracted signal suppressing with a uniform step these points. In this way, we are certain that we act only on the portions of the signal far from peak (that do not correspond to significant reflections such as Back echo reflection, Frontal reflection, and the reflections caused by defects) and that the suppression is uniform in all the portions of non-significant signal. In Fig. 2, the normalization procedure is shown. In the first row, the original signal is shown, then in the second row the signal obtained considering only the points that are under a fixed threshold is reported. On this signal, a number of points are suppressed with a uniform step. Finally, in the last row the reduced signal is reported. It still maintains the same shape and the peak preserves the same relative position along the signal. In this particular case, the normalization consisted of the suppression of 22 samples. The effectiveness of the proposed techniques has been demonstrated by the final classification results that are better than those obtained with other interpolation techniques. 4. Defect segmentation and classification After the normalization step, the ultrasonic signals are ready to be processed for defect detection and recognition. We used a two-levels approach for the detection and the recognition phases. Firstly, the normalized ultrasonic signals are processed to separate defective from sound areas (segmentation process), then the resulting segmented image is analyzed to classify the defective areas on the basis of their defect type and localization in plies (classification process). In this paper, both the segmentation and the classification steps are performed by a neural network. Neural networks are applicable because they have a natural propensity for storing experiential knowledge and making

4 148 ARTICLE IN PRESS T. D Orazio et al. / NDT&E International 41 (2008) Fig. 2. The ultrasonic signals corresponding to the original signal (top), the original signal that is under a threshold with suppressed samples marked in red (middle), and the resulting normalized signal (bottom). it available for later use. The knowledge is acquired by the network through a learning process that varies the interneuron connection strengths, known as synaptic weights [8]. The great success of neural networks derives from their ability to solve complex problems that require a non-linear mapping between the input space and the output space, and their generalization capability, which allows them to produce reasonable outputs from inputs that were not encountered during learning. In this paper, we have used a three-layer neural network, characterized by the presence of an input layer of source nodes one hidden layer, and an output layer. The number of neurons in the input equals to the number of signal features. Since we provide the normalized signals to the network, the number of input nodes is equal to the number of sample points extracted from the resulting continuous signal. The number of nodes in the output layer depends on the number of classes that the network has to recognize. The number of nodes in the hidden layer is determined by the experiment. The hidden layer enables the network to extract higher-order statistics especially when the number of the input layer is high. A supervised learning approach has to modify the synaptic weights connecting the neurons of the network. For each network, a set of training examples is fixed and the training is repeated using different configurations of the hidden layer. The set of training examples consists of input output couples (input signal, corresponding desired response). Some points of known examples are extracted from the considered materials and continuously fed into the net so that the synaptic weights are tuned to ensure the minimum distance between the actual and the desired output of the net. Training continues until a steady state is reached, i.e., no further significant change in the synaptic weights could be made to improve net performance. The first step of defect segmentation is performed by using a neural network with two output neurons. Each available normalized ultrasonic signal is fed into the net, which classifies it as either relative to defective areas or a sound area. According to the neural network outputs, a binary image is produced containing black points for defective areas and white points for sound areas. Only the signals classified as belonging to defective areas are fed as input to the classification step. The classification step classifies defect type, defect position or both using three different neural networks. The network for the classification of defect types has three output nodes corresponding to the three different defect types under consideration in our experiments. The network for the classification of defect position has three output nodes corresponding to the Bottom, Medium, and Top positions of defects. The last network for the simultaneous classification of defect type and position has nine output nodes corresponding to all the possible combinations of type and position. 5. Experimental results Experimental tests were performed on three different composite materials (referred to as standard in the following) with a honeycomb structure with Nomex Core and different thicknesses (48, 64, and 128 plies where each ply has a depth of 0.19 mm). Ultrasonic data were obtained by an ultrasonic reflection technique that uses a single transducer serving as transmitter and receiver (5 MHz). Each standard contains artificial defects introduced during the manufacturing process and composed of the following materials: Brass Foil, Pressure Sensitive Tape, Dry Peel Ply (in the following, [A] stands for Tape, [F] for Peel Ply, [B] for Brass). The length of the signals extracted from the three different materials is 77, 99, and 183 points, respectively. For each standard, defects were positioned as reported in Fig. 3. The typical insert locations are: (1) two plies from toolside surface for Brass and Pressure Sensitive Tape and five plies for Peel Ply (top), (2) mid-part thickness (mid), and

5 ARTICLE IN PRESS T. D Orazio et al. / NDT&E International 41 (2008) A-b A-b B-b A-b B-b B-b F-b F-b [A] TAPE A-m B-m A-m B-m B-m A-m F-m F-m [F] PEEL PLY [B] BRASS A-t A-t B-t A-t B-t B-t F-t F-t t = top m = middle b = bottom Fig. 3. Defects types and locations applicable on each reference standard. Sound Area A-m defect (Tape-middle) 48 plies 64 plies 128 plies Fig. 4. The A-scan presentations of background and A-m defect points of the three standards. The x-axis represents the time and the y-axis the energy of the ultrasonic signal. (3) one and two plies from bagside surface for Brass and Pressure Sensitive Tape and five plies from bagside surface for Peel Ply (bottom). In Fig. 4, the A-scan presentations of some background and defect points are displayed the images on the left show the continuous signals generated from points selected in sound areas and on the right, the continuous signal generated from points in defect areas. For materials with 48 plies, it is possible to see also successive peaks corresponding to the reflection of the ultrasonic impulses. As explained above, we have considered only the portion of the signal between the main Frontal and Back echoes. The normalization procedure eliminates 22 samples to reduce the signals coming from the 64 plies material (99 77 points) and 106 samples to reduce the signals coming from the most thick material with 128 plies ( points). Experiments were performed by using a different number of neurons in the hidden layer in order to discover the optimum for the application domain under consideration. The training of the neural network was performed using the well-known Back Propagation algorithm [9]. In particular, for segmentation the network was trained using

6 150 ARTICLE IN PRESS T. D Orazio et al. / NDT&E International 41 (2008) signals extracted from the thinnest standard (48 plies). These signals consisted of different groups of points manually selected from different classes of defects and locations (100 points from background, 15 points from A-t, A-m, A-b, B-t, B-m, B-b defects, and 10 points from F-t, F-m, F-b). In Fig. 5, it is possible to observe the C-scan representation of the standards used to train the net for the segmentation phase. White color represents the points used in the training phase. In the classification phase, the three neural networks were trained using 120 signals extracted from the thinnest standard (48 plies). The selected points are the same as those used in the training of the segmentation phase but excluding the points corresponding to sounds areas. Fig. 5. The points selected in order to train the network used in the segmentation phase. The following subsections detail the experimental results on the three subject standards for both defect detection (Section 5.1) and classification (Sections 5.2, 5.3, and 5.4). To evaluate experimental results, a quantitative analysis was performed using scatter matrices in which ground truth data were pointwise compared with the output of the neural networks Segmentation of defective areas Fig. 6 shows the three black-and-white images obtained after the application of the segmentation phase described in Section 4. Black pixels are associated to defective samples whereas white pixels are associated to sound samples. Table 1 reports the relative scatter matrix for a quantitative evaluation of the segmentation performance of the system. The results show that the neural network separates very well the defective areas from sound areas in two of the three standards. The percentages of the correctly detected defective points were greater than 97% for the first two standards and about 55% for the third standard. In the latter, only one defective area was completely missed in the standard with thickness of 128 plies, the system did not detect the defect B placed on the Bottom (B-b defect). We believe that the reason the detection performance worsened for the thickest standard was the progressive reduction of signal-to-noise ratio. In fact, during the propagation of the ultrasonic signals through the internal component of the standard, a noise component increasingly affects the echoes. If the thickness of the inspected component exceeds a certain number of plies (depending on the properties of the inspected component), the noise modifies the shape of the ultrasonic echoes making the inspection less efficacious. It was not coincidental that the missed detection which occurred in our experiment was owing to this unpleasant effect of increasing noise in correspondence to a Brass defect. In general, ultrasonic Brass defect echoes were the most similar to the sound echoes. When the Brass defects were placed on the Bottom of the standard, the noise component altered the shape of the signal substantially rendering it indistinguishable from the sound echoes. Defect segmentation 48 plies Standard 1 64 plies Standard 128 plies Standard Fig. 6. The segmented images produced by the neural network.

7 ARTICLE IN PRESS T. D Orazio et al. / NDT&E International 41 (2008) Table 1 In the first column the experimental ground truth is reported Real defect Out of NN 48 plies 64 plies 128 plies Background Defect Background Defect Background Defect Background (22434) (98.28%) 385 (1.71%) (97.67%) 522 (2.32%) (98.49%) 338 (1.50%) Defect A-t (150) 1 (0.66%) 149 (99.33%) 9 (6%) 141 (94%) 24 (16%) 126 (84%) Defect A-m (150) 0 (0%) 150 (100%) 4 (2.66%) 146 (97.33%) 9 (6%) 141 (94%) Defect A-b (150) 2 (1.33%) 148 (98.66%) 6 (4%) 144 (96%) 5 (3.33%) 145 (96.66%) Defect B-t (150) 1 (0.66%) 149 (99.33%) 4 (2.66%) 146 (97.33%) 24 (16%) 126 (84%) Defect B-m (150) 1 (0.66%) 149 (99.33%) 29 (19.33%) 121 (80.66%) 68 (45.33%) 82 (54.66%) Defect B-b (150) 24 (16%) 126 (84%) 24 (16%) 126 (84%) 150 (100%) 0 (0%) Defect F-t (289) 4 (1.38%) 285 (98.61%) 53 (18.33%) 236 (81.66%) 4 (1.38%) 285 (98.61%) Defect F-m (289) 1 (0.34%) 288 (99.65%) 23 (7.95%) 266 (92.04%) 51 (17.64%) 238 (82.35%) Defect F-b (289) 0 (0%) 289 (100%) 0 (0%) 289 (100%) 128 (44%) 161 (55.70%) In the remaining columns the segmentation outcomes of the proposed system are reported. Table 2 The defect position classification results for the three composite materials (48, 64, and 128 plies) Defect type Standard (plies) (# points) Top (%) Middle (%) Bottom (%) A-t 48 (149) (141) (126) 100 A-m 48 (150) (146) (141) 96 4 A-b 48 (148) (144) (145) 100 B-t 48 (149) (146) (126) 100 B-m 48 (149) (121) (82) 96 4 B-b 48 (126) (126) (0) F-t 48 (285) (236) (285) 99 1 F-m 48 (288) (266) (238) 94 6 F-b 48 (289) (289) (161) 100 Table 3 The defect type classification results for the three composite materials (48, 64, and 128 plies) Defect type Standard (plies) (# points) A (%) B (%) F (%) A-t 48 (149) (141) (126) 95 5 A-m 48 (150) (146) (141) A-b 48 (148) (144) (145) B-t 48 (149) (146) (126) B-m 48 (149) (121) (82) B-b 48 (126) (126) (0) F-t 48 (285) (236) (285) 100 F-m 48 (288) (266) (238) F-b 48 (289) (289) (161) 100 A possible solution for this kind of problem is to increase the frequency of the inspection signals although this would introduce a potential negative effect since the more superficial defects may be lost. By the same reasoning, the false defective points scattered across the deepest standards could be avoided. However, this problem is not

8 152 ARTICLE IN PRESS T. D Orazio et al. / NDT&E International 41 (2008) significant since that the percentage of false detection was never greater than 3% and it did not alter substantially the qualitative perception of the defective areas in Fig Identification of defect position In this experiment, the neural network has been trained, as described in Section 4, in order to identify the ply position of the defective points found in the segmentation phase. In Table 2, the classification results are reported for the three composite materials. In the first and second columns, the defect type and the number of test points for each standard are reported. The corresponding classification percentages are recorded in the Top, Middle, and Bottom columns. For A-t, B-t, F-t defect types, the correct detections are expected in the Top column; for A-m, B-m, F-m defect types in the Middle column; and for A-b, B-b, F-b defect types in the Bottom column. It should be noted that the B-b defect in the 128 plies standard has not been considered since it was not detected in the first step of defect segmentation. The results are very encouraging for the three considered standards the percentage of points correctly classified is 100% for the 48 plies standard, 99% for the 64 plies standard, and 94% for the 128 plies standard. The classification performance decreased when the thickness of considered standards increased and this is a consequence of the signal/noise ratio effects reported in the previous subsection Identification of defect type In this experiment, the neural network was trained in order to identify the type of defect. Table 3 reports the quantitative results on the three standards. The neural network properly classified all the defective regions except for the regions of F-m defect in the two thickest standards. In fact, in the 64 and 128 plies standards most of the points in these regions are classified as B-m defect. This drawback was caused by both the noise introduced from internal echoes and the similarity between ultrasonic echoes from the defective areas under consideration. Summing up, the classification performance for the 48 plies standard was satisfying apart from the F-m defect point evaluation, where only 63% of points were correctly classified. For the 64 plies standard, the neural network classified correctly most of the defective areas (93% of correct classification). The percentage of correct classification decreased for A-m, B-m, and F-m defects (43%, 67%, and 45%, respectively). For the 128 plies standard, the neural network failed to optimally classify only A-m, A-b, and F-m defects (percentages of correct recognition were 56%, 66%, and 50%, respectively). Table 4 The defect-type and position classification results for the three composite materials (48, 64, and 128 plies) Defect type Standard (plies) (# points) A-t (%) A-m (%) A-b (%) B-t (%) B-m (%) B-b (%) F-t (%) F-m (%) F-b (%) A-t 48 (149) (141) (126) A-m 48 (150) (146) (141) A-b 48 (148) (144) (145) B-t 48 (149) (146) (126) B-m 48 (149) (121) (82) B-b 48 (126) (126) (0) F-t 48 (285) (236) (285) F-m 48 (288) (266) (238) F-b 48 (289) (289) (161)

9 ARTICLE IN PRESS T. D Orazio et al. / NDT&E International 41 (2008) Identification of defect type and position In this experiment, the neural network was trained to identify the type and the plies position of defect points both at the same time. Table 4 shows in detail the experimental results on the three standards. This task is more difficult than the previous ones. The performance is optimal for the 48 plies standard, where the lowest percentage of correct recognition is 78% (for the classification of F-m defect area). In the case of 64 plies standard, the performance is over 95% except for A-m, B-m, and F-m defects, where the percentage of points correctly recognized is 54%, 71%, and 47%, respectively. The performance on the 128 plies standard is inferior to those on the other standards as with the other experiments. In fact, the neural network fails completely to classify the points of A-m and F-m defects, where the percentage of correct recognition is 14% and 37%, respectively. However, for all the other defective regions of the 128 plies standard, the percentage of the points correctly recognized is over 50%. The best performances (90% of points correctly recognized) are obtained when the neural network analyzed the A-t and F-b areas. Within our experimented settings, the worst performance occurred during the analysis of the 128 plies standard. We assume this occurs because the difference in thickness between the tested standard (128 plies) and the training example (48 plies) is experimentally relevant, producing a notable difference between the signals produced. We use our algorithm of normalization to place the Back echo reflection and the reflections caused by defect in the same positions along the horizontal axis that are occupied by the signals from the 48 plies standard. The position along the horizontal axis is the important information for recognizing the ply position of defects. Therefore, our procedure operates well also in the case of 128 plies standard. In fact, the performance in the experiment to identify the defect ply position (detailed in Table 2) is over 94%. Unfortunately, the performance decreases when we classify the type of defects in the 128 plies standard because the shape of the reflections from the defects is not totally equal to that of the 48 plies standard due to the noise occurring through composite materials during the acquisition of ultrasonic data. Furthermore, the results obtained with a two-step classification (two different networks for position and type classification) are better than those obtained with a singlestep classification, also for the 128 plies standard, and are encouraging for building an automatic system for internal defect detection whatever the thickness may be of the composite material. 6. Conclusion and future works A large amount of research work has been conducted using various NDT&E techniques in the detection and identification of defects in composite materials. Ultrasonic inspection has proven to be effective in the assessment of internal defects. In this paper, we have presented a neuralnetwork-based analysis of ultrasonic data for the automatic detection of internal defects in three different composite materials with different thicknesses. Among supervised approaches, neural networks have been preferred since they require a simple training phase once at the beginning and then they can be quickly applied during the test phase. The system requires only the selection of a set of training examples extracted from defective areas and a sound area from a single standard. The neural network is able to build an internal representation to map the input signals to the desired output and is able to generalize to new and different areas of materials not encountered in the training phase. The main result of this work is the possibility of using a neural network trained only on one kind of material and then, by using a normalization procedure, applying the same network to different thickness materials. Results from experiments carried out on three different composite materials demonstrate the effectiveness of the proposed approach. In particular, the network is always able to recognize the defect positions on the three standards, and the defect type when positioned on the Bottom and on the Top plies of the materials. For the medium position, the performance of defect-type recognition decreases since the signals produced by the Brass and Peel ply defects are very similar. In this case, the system is able to correctly detect the defect positions but there is a confusion between two different type characterizations. The main limit of the proposed approach is in the requirement of a set of positive and negative examples for training the neural network. In our opinion, this constraint is not so strict since the usage of a reference material containing the possible defects at different positions is a well-established procedure normally used to validate the acquisition done with any sensor. Processing times are very encouraging for the application of the proposed algorithm during real maintenance controls the training phase of the neural networks, that requires some minutes to converge to the optimal configuration, is carried out just once at the beginning using the standard reference material, then the test phase requires just few seconds to process a matrix of ( ) where ( ) is the image dimension and 77 is the signal length. Considering that algorithms have been developed in Matlab, further improvements in time processing can be achieved converting the algorithms in C++ and applying code optimization. Future work will be addressed to the analysis of the connectivity of defect regions to establish the correct defect-type association according to the coherence of network results in the region. References [1] Huang YD, Froyen L, Wevers M. Quality control and non-destructive test in metal matrix composites. J Nondestr Eval 2001;20(3):

10 154 ARTICLE IN PRESS T. D Orazio et al. / NDT&E International 41 (2008) [2] Edwards RS, Dixon S, Jian X. Characterisation of defects in the railhead using ultrasonic surface waves. NDT&E Int 2006;39: [3] Ruiz-Reyes N, Vera-Candeas P, Curpia n-alonso J, Cuevas-Martı nez JC, Blanco-Claraco JL. High-resolution pursuit for detecting flaw echoes close to the material surface in ultrasonic NDT. NDT&E Int 2006;39: [4] Ravanbod H. Application of neuro-fuzzy techniques in oil pipeline ultrasonic nondestructive testing. NDT&E Int 2005;38: [5] Bettayeb F, Rachedi T, Benbartaoui H. An improved automated ultrasonic NDE system by wavelet and neuron networks. Ultrasonics 2004;42(1 9): [6] Gil Pita R, Vicen R, Rosa M, Jarabo MP, Vera P, Curpian J. Ultrasonic flaw detection using radial basis function networks (RBFNs). Ultrasonics 2004;42: [7] Margrave FW, Rigas K, Bradley DA, Barrowcliffe P. The use of neural network in ultrasonic flaw detection. Measurement 1999;25: [8] Haykin S. Neural network a comprehensive foundation. New York: IEEE Press; [9] Freeman J, Skapura D. Neural network algorithms, applications, and programming techniques. Reading, MA: Addison-Wesley; 1991.