Multilayered composite plates are widely used in many advanced industries for their properties. One of the pivotal problems still remains efficient detection and characterisation of delaminations occurring in multilayered thin composite plates, the thickness of which is less than the spatial length of the ultrasonic pulse. Limited bandwidth of a commonly used ultrasonic testing equipment distorts the information about delaminations so these are not easily detectable. Two different techniques have been developed for detection and characterisation of the delaminations in thin multilayered composite materials using band-limited ultrasonic signals acquired by commercially available instrumentation.
A reliable method for delamination inspection of thin laminates was developed, based on differential signal extraction (DSE). The major achievement of this method is that time domain information is sustained: the method can resolve the layer at which the delamination has occurred.
All the factors mentioned degrade the measurement results.
Usually, results are presented as 1-, 2- or 3-dimensional images.
Therefore it is frequently said that image quality is corrupted or blurred. All the algorithms mentioned are dedicated to increasing image quality which has been degraded by all the factors mentioned. There is no unique algorithm, capable of fighting everything at once. Every algorithm is dedicated to solving one of the problems. Sometimes it is even necessary to sacrifice some signal properties in order to gain others. Here is the list of known image processing methods investigated and employed in this thesis: Inverse filtering, Wiener filtering, Autoregressive spectral extrapolation (ASE), Split spectrum processing, Synthetic aperture focusing technique (SAFT), and Ultrasonic spectroscopy.
|It is always desirable to perform the measurements with the best possible accuracy, but in order to do this, additional information must be obtained. In ultrasonic NDT this is done by initial measurements. During the measurement process, optimal system gain is of major consideration. The method we propose allows us to obtain the data without initial measurement- "in one shot". This greatly increases system speed and improves the accuracy of measurements. Comparison with other popular methods is done: Amplifier gain adjustment during multiple scans, Iterative gain adjustment, Predictive gain adjustment. We have developed predictive gain adjustment, which is analysed here. In this algorithm the gain of the system is adjusted during the scanning process. At each position of the transducer, the previous values of the MaxPlot signals are used to predict the signal level one scanning step forward, and consequently to set the necessary gain factor of the amplifier. It allows collection of the signals required to produce one B-scan during a single scan of the transducer.
After A/D conversion, low-pass filtering is used to reduce the influence of noise. An exponential second order filter is used: |
, ( 2.1)
, ( 2.2)
here UI and UII are the signals corresponding to the output of the first and second stages of the filter, xi is the ith position, and tf is the filter constant (<1). Higher order filters would give better noise reduction but might increase the computation time significantly. The prediction block extrapolates the signal values one scanning step. For this purpose the three point Lagrange polynomial extrapolation is used:
, ( 2.3)
here UA/D(xi+1) is the signal level at the amplifier input predicted one step forward xi+1. Such a simple expression can be used thanks to the constant scanning step. The gain control unit calculates the signal level at the A/D converter input according to the current gain value,
, ( 2.4)
and decides whether such gain will be suitable, or if gain correction is required. The correction is performed if the predicted signal exceeds the allowed limits Ulmin and Ulmax:
The gain correction block adjusts the current gain according to the predicted signal value and the present limits. All gain factor values K(xi) are stored in the computer's memory, so the signal level at the amplifier input can be restored:
, ( 2.5)
where uA/D(xi,zk) is the complete signal waveform in the time domain at the xi scanning position at the A/D converter output. The waveform at the input u(xi,zk) restoration is equivalent to an expansion of the system's dynamic range, because after such an operation the influence of the gain K(xi) is eliminated and we again have data similar to that before the range compression. A restored signal has a wider range, as if it were stored only in an A/D converter representation. This process is illustrated by corresponding time diagrams in Figure 2.1
Figure 2.1. Time diagrams illustrating the adaptive gain adjustment (SNR=80dB): gain of the controlled amplifier (top); actual (solid line) and predicted (grey line) signal amplitudes at the amplifier output (central); restored MaxPlot signal (bottom).
The accuracy can be estimated analytically or by means of a numerical simulation. The calculation time can be determined only by computer modelling. Analytical calculation resultsfor 6 and 8 bit A/D converters are shown in Figure 2.2. These estimations take into account only quantisation errors. There are other sources of errors in acoustic imaging systems, such as acoustic and electronic noise, a finite dynamic range of the controlled amplifier, etc. The evaluation of the complete error can be performed by a computer simulation using the modelling program described above. It is necessary, however, to find a criterion by which it would be possible to evaluate the integral relative error value of the small as well as large amplitude MaxPlot signals obtained with different control algorithms. The well-known root mean square estimation does not fulfill this requirement. Therefore, we propose to evaluate this error by absolute deviation normalised by the current values of the signals at each scanning point; this definition can be called L1 norm: |
, ( 2.6)
where URi is the actual value of the signal given in a floating point representation, UMi is the nA/D bits representation of the signal after A/D conversion, and n is the number of measurements performed. Such an evaluation of the error allows us to see the advantages of the adaptive gain adjustment. The efficiency of the different gain control algorithms can also be evaluated assessing the time needed for the calculations (presented in thesis).
Figure 2.2 Relative errors of different control algorithms: iterative gain adjustment; adaptive gain adjustment with prediction; gain adjustment without prediction; logarithmic amplifier. Index 8 stands for 8 bit A/D converter, 6 - the 6 bit A/D.
The application of the algorithm on data acquisition showed good performance of adaptive gain control. Great improvement of the images is obtained if adaptive gain control is used in the case of a relatively high SNR. This is illustrated in Figure 2.3 and Figure 2.4 by the images of a point-type reflector, located beside a large, flat reflector, obtained by adaptive and conventional acquisitions.
Image (SNR=60dB) restored by SAFT after the adaptive data acquisition: MaxPlot image (left), B-scan image (right)
Point-type reflector image restored by SAFT after conventional data acquisition: MaxPlot image (left), B-scan image (right).
We can see that adaptively-coded data yield better results. The differences between conventional and adaptively acquired images increase in the case of A/D converters with a smaller number of bits. An adaptive imaging algorithm for ultrasonic NDT purposes results in significant compression of the dynamic range of the received ultrasonic signal, so high quality acoustic images can be obtained even using 6-8 bit high speed A/D converters. The hardware needed to realise this method includes an amplifier with a computer controlled gain factor. All other operations are performed by software. The simplicity of the method described can be used to modify any conventional 2D acoustic imaging system. This was implemented on the IZOGRAF-3 acoustic imaging system.
The main advantage of the technique proposed is the possibility of obtaining a B-scan image during a single scan of a transducer without a priori detailed knowledge about the expected range of ultrasonic signal amplitudes. This feature is of great importance when NDT is performed in a dangerous environment. It is necessary to point out that it is similarly possible to compress the dynamic range of C-scan images.
The compression of dynamic range is performed without a loss of information, becausethey are represented independently of the absolute level of the input signals in the buffer memory with the same relative accuracy. The actual signal level can be restored from the gain factors stored in computer memory. This is equivalent to the digitisation of the signal with more bits than are available in the A/D converter used. Therefore, this algorithm has additional advantages in the case of advanced acoustic image post-processing.
There is also the problem of high structural noise due to scattering from the fibres, even at the lower frequencies. In this thesis we report the application of the well-known split-spectrum processing method (SSP) to the anisotropic composite and show that it reduces much of the structural noise. Improvement of the axial resolution is demonstrated using a novel application of Wiener filtering and Autoregressive Spectral Extrapolation (ASE). Both these techniques applied in the conventional sense would be unsatisfactory in the case of this high loss material, because of the difficulty in obtaining a suitable reference 'wavelet'. With a new approach to attenuation compensation and judicious application of Wiener filters to 'split-frequency' signals, we will demonstrate not only an improvement in the axial resolution but also in the structural noise level in images of defects in this composite material.
The fragment of an X-radiograph of the test specimen is shown in Figure 3.1. a. This specimen is an aluminium-6082 matrix composite with 43% volume fraction of low modulus, long carbon fibre reinforcements in a [0°,90°] lay-up. The thickness of the plate is 12mm. The flat-bottomed holes have three different diameters and are at four different depths from the surface. These holes were used for establishing the performance of the imaging system, before and after signal-processing. The figure demonstrates the 10mm diameter, 10mm depth hole. Figure 3.1 b presents the 3-dimensional image obtained after the new processing scheme.
The X-ray (left zoom) of an Al-C metal-matrix composite sample and the output image of the new processing method ( right zoom).
Before the holes were imaged, it was necessary to obtain the attenuation versus frequency characteristics of the specimen. The usual method for measuring attenuation in a plate is to look at the ratio of consecutive backwall echoes. But since the attenuation in this specimen is very high, only one backwall reflection is measurable. Therefore a special through-transmission arrangement was employed which is explained in the thesis. The attenuation information a(f) was used to choose the best transducer frequency range within which the images were obtained.
Two improvements to the signal quality are needed, before the B-scan or C-scan image is constructed from the processed A-scans. The first is an SNR enhancement to be achieved by split-spectrum processing (SSP), and the second is a depth resolution enhancement to be obtained by Wiener filtering (simplest case). The effectiveness of Wiener filtering depends on the correct choice of the reference wavelet. This is also true for ASE because it uses a Wiener filter as the first stage. For high loss materials no suitable reference wavelet can be found, as explained earlier. A new approach is presented here whereby loss compensation is applied to the signals in such a way as to make the use of Wiener filtering more effective, at the same time delivering all the benefits of SNR enhancement obtained with SSP.
The inputs to the process are the reflected signal s(ti) and a reference wavelet r(t). This reference wavelet could be the front face echo from any location on the sample. First both inputs are Fourier transformed into the frequency domain. Then each signal is split up by multiplication of each spectrum with the bank of Gaussian windows. The process of 'frequency-selective loss compensation' (FSLC) is applied next. The split-frequency windows are inverse Fourier transformed to yield a set of n elements of time domain signals swj(ti):
here F and F-1 stands for forward and back Fourier transform, Wj is Gaussian window:
thus w j being the central frequency and bwj - bandwidth of jth window. The same way the reference wavelet is split into windows:
Each split signal is then multiplied by an exponential gain compensation term:
where a j is the attenuation coefficient at the centre frequency of the jth Gaussian. The values used are taken from the experimental a(f). The reference wavelet is transformed the same way:
Transformation of the signals to the frequency domain by FFT completes the compensation process. With both the split signal and the split reference wavelet correctly compensated for the effect of frequency dependent attenuation, a Wiener filter is applied to each split signal before the resulting deconvolved signal is transformed back into the time domain.
This leaves a set of split time domain signals sdj(t), each comprised of a series of narrow spikes. These split signals still contain both structural noise and any processing noise generated by the Wiener filter, however. The final stage, the nonlinear processing of these split signals, is performed by maximisation followed by polarity thresholding, removing most of the structural and filter generated noise:
spt is an output signal at the time instant ti. Averaging and polarity is performed in the same manner. The FSLC process is expected to make the performance of the Wiener filter less sensitive to the choice of the reference wavelet. This should enable one to use the front face echo as the reference r(t) and still obtain a sharpening of a deep defect echo.
In order to prove the stability of the FSLC method and have a better presentation of what the defect looks like, 4-dimensional data was collected (x,y,t and amplitude). Using a standard -6dB threshold, a 3-dimensional isosurface of the defect image was created (Figure 3.1 b). The image correlates with data obtained using X-rays of computed tomography (results presented in thesis).
A new signal processing scheme for ultrasonic imaging of high loss composite structures incorporates FSLC and Wiener filtering within a split-spectrum processing algorithm. It was shown that this approach greatly enhances the SNR and improves the depth resolution for images obtained in a specimen of high acoustic noise level. In comparison with other signal processing methods it showed the best performance. The good performance of the algorithm can be explained by time-frequency dependent signal components and proper correction by FSLC, both of which contribute to good SSP algorithm performance. Split windows contribute to a better fit of the reference wavelet and signal in inverse filtering. All the factors mentioned combine to produce an amazing performance. The results inquiry with X-ray CT gave a good match on results.
We introduce signal processing algorithms enabling efficient detection and characterisation of delaminations in thin multilayered composites using band-limited ultrasonic signals acquired by commonly used ultrasonic instrumentation. Particularly, the algorithms exploiting the modified L1 norm deconvolution and frequency domain imaging are analyzed. The algorithms were tested both on simulated and real signals. For this purpose a simplified 1D mathematical model based on a 4-pole matrix approach was developed and experimentally verified. Comparison of the simulated and experimentally obtained results revealed the possibility of determining the character of the delamination occurring in layered laminates.
|The specimen used for this purpose was an ARALL type 3 layer composite plate with an unidirectional glass fiber reinforced prepreg between two aluminum 7075-T6 layers. The acoustical and geometrical parameters correspond to those used in calculations. The composite plate contained an artificial delamination which arose during a fatigue test around the hole drilled through the specimen (Figure 4.1).|
Figure 4.1. Sample "view" (zoom left),
corresponding through transmission image (zoom right)
All experiments were carried out using the ultrasonic imaging system IZOGRAF-2 in pulse-echo mode with 20 MHz sampling frequency and 8 bit data conversion. The composite plate was immersed in water and located 40 mm from the 5 MHz ultrasonic focused transducer corresponding to its focal distance. Ultrasonic data acquisition was performed from two opposite sides of the plate in order to reveal the differences in the obtained images. The images cover a region 40x40 mm around the hole drilled in the center of the plate. In order to prevent water penetration inside the delamination at the vicinity of the hole, the crack opening was covered by Scotch tape stuck to a surface of the sample. For reduction of the influence of electrical and quantisation noise received pulses were averaged.
The image of the same sample but produced using through transmission technique is displayed in Figure 4.1 right. The dark object in the center represents the delamination lateral configuration. This result can be trusted, because through-transmission is a reliable method, widely used in practice. Also, a 15mhz frequency was used, which is 3 times the frequency we are planning to use for our investigation. Also, an 0.1mm step was used to acquire the image. The rest of the investigation was performed with a 0.6-1.2mm step. The raw image for pulse-echo investigation has a rather low signal/noise ratio and does not allow us to determine reliably either the geometrical shape or the depth of the delaminations.
|The well known L1 norm processing was modified in order to extract necessary features from the data. The results presented indicate that the proposed filtering and image formation procedures enhance the images significantly. As should be expected, C-Integral scan images possess better SNR than C-MaxPlot scan images. After compression of the filtered A-scans into one point of the 2D image the spatial resolution along z-axis is lost, but still enables us to see differences between delaminations on the opposite sides of the hole. That can be achieved by collecting data from two opposite sides of the plate and displaying the obtained filtered C-type images. For example, the top view displays the flaw only on the left side (Figure 4.2 left) and the bottom view only on the right side of the hole (Figure 4.2 right), which allows us to suggest that they are located between different layers. In other words, disruption has occurred between different layers on the left and the right side of the hole (Figure 4.3). The image shown in Figure 4.2 is obtained using a modified algorithm, which exploits the residual part of L1 norm deconvolution's first iteration.|
C-Integral image after iterative front face subtraction.
Scan from "top"(zoom left)
and "bottom" (zoom right) side.
A similar result to L1 norm deconvolution can be obtained by displaying not the differential pulse response, but the residual part of the reflected signal , which is determined by subtracting from the reference signal , reflected by a flawless region, the pulse , received at the delaminated area. The reflection from the flawless region was used here to deconvolve with instead of the reference wavelet. In such a setup the residual part even gains the differences between the delaminated and flawless areas.
As it follows from a variety of publications and also from the results of the simulations, spectra of the pulses reflected from perfectly bonded and delaminated regions are different, what makes their characterisation feasible. However, in the case of band-limited signals and very thin samples such a characterization is complicated by a strong frontal echo which masks an interference pattern caused by disbonds. The proposed L1 norm deconvolution procedure's residual use enables us to overcome this shortcoming. The FFT is performed not on the complete received signal, but on the above mentioned residual part of each received signal :
There F stands for the Fourier transformation. That gives the spectrum only of the internal reflection, containing mainly information about the intrinsic structure of the composite. The normalized magnitudes of the selected frequency are displayed as a 2D image
After such a transformation a spatial resolution along z axis is lost, but by presenting images at various frequencies it is possible to govern a spatial resolution in x,y plane and to extract different information about disbonds located at different distances from a sample's surface. That allows us to obtain both sides' delamination images from one side.
| Figure 4.3.|
C-scan of frequency domain at 3.75mhz after iterative front face subtraction (left). Right picture presents the interpretation of investigation results.
Both techniques, especially the second one, work as feature extraction procedures, revealing differences between pulses reflected by flawless and delaminated regions and in our opinion are not limited to the case of 3 layer composite plates.
An adaptive imaging algorithm for ultrasonic NDT purposes has been developed. It results in significant compression of the dynamic range of the received ultrasonic signal. The main advantage of the technique proposed is the possibility of obtaining a B-scan image during a single scan of a transducer without a priori detailed knowledge about the expected range of ultrasonic signal amplitudes. This feature is of great importance when NDT is performed in a dangerous environment. The algorithm has additional advantages in the case of advanced acoustic image post-processing, for instance SAFT.
The Paper was presented on the UTonline Application Workshop in May '97
|| UTonline ||