|NDT.net May 2004 Vol. 9 No.05|
| CT-IP 2003 Proceedings
Automatic delamination defects detection in radiographic sequences of rocket boostersRebuffel V. and Pires S. (CEA-LETI/DSIS, 17 rue des Martyrs, F38054 Grenoble cedex 9,
France, e-mail : firstname.lastname@example.org),
Caplier A. (INPG, Grenoble, France), Lamarque P. (SNPE, St Mיdard en Jalles, France)
AbstractSolid rocket motors are routinely examined in real-time X-ray radioscopic mode. In a tangential configuration, the peripheral part of the object is imaged in order to detect the delamination defects between the propellant and the external metal envelope. We have previously developed a real-time imaging system, based on a tomosynthesis algorithm in order to improve the contrast. In this paper, we present a method to perform the automatic detection of delaminations. A single image being not sufficient for robustness, a spatio-temporal aspect is required for the algorithm. In a first step, the proposed method computes the displacement between the current radiograph and a reference one. The images are then registered and compared. On the resulting difference image we apply a smoothed thresholding to obtain an instantaneous confidence image. These images are then cumulated over time to get maps of probability of a defect pixel. Defects are then enhanced thanks to a neighbourhood validation. Thus the method combines spatial and temporal aspects in both detection and characterization. Several experimental tests have been performed, on both true and simulated radiographs of boosters and of other cylindrical objects.
1. IntroductionThe traditional X-ray inspection technique for the examination of large objects works in a radioscopic mode : the object intersects the X-ray flux, and moves so that successive parts of it can be visualized. If the motion of the object is low, a strong correlation exists between successive views, and potential defects may be visible over several radiographs. When using a digital system, the obtained sequence of radiographs can be analysed by an operator, but also can be combined numerically, allowing a more reliable and quantitative control.
This configuration is used for the examination of the solid motors of Ariane V rocket. The large and cylindrical boosters are rotating between a high energy X-ray source and a 2D X-ray detector. The system should process in real time during the rotation of the booster. The purpose of the control is to detect possible defects all through the sample with contrast down to 2%. In the tangential configuration, the part of the object that intersects the X-ray beam is the peripheral one, allowing to detect the delamination defects between the propellant and the external metal envelope. In the opposite, the radial configuration focuses on air cavities in the mass of propellant. In this study we only consider the tangential configuration.
Defect detectability is very poor due to the strong attenuation of the high energy X-rays through the motors. We have previously developed a real-time imaging system, processing the radiographs on line, and based on a tomosynthesis reconstruction algorithm in order to improve the signal-to-noise ratio [ANTO01]. Without tomosynthesis process, small defects are hardly visible. After tomosynthesis, their detection is easier but still difficult, in the sense that a single image is often not sufficient to conclude on the presence of the defect. However, an operator also needs a number of successive images to be able to status on the presence of a defect.
Our purpose is to specify and implement an algorithm processing the sequence of radiographs to extract and characterize the possible defects. Due to the poor detectability previously mentioned, this algorithm requires a spatio-temporal aspect to achieve robustness. It should be able to process a sequence of radiographs with or without preliminary tomosynthesis processing, the results being obviously better in the first case. In the following, first we present the industrial system that has been implemented for the X-ray inspection of the Ariane booster in tangential configuration, and describe the tomosynthesis process. Then we introduce our method, and detail each step. Experimental results are shown in the last section.
2. The industrial system used for X-ray inspection of the boosters
2.1 Description of the systemThe ARIANE V solid motors are made of 3 meters diameter segments. Each segment contains 110 tons of solid propellant in a cylindrical metal envelope (thickness 20 mm) on which is laid a thermal protection. A central channel allows the ignition and the combustion gas draining. The control requirements are the detection of air cavities in the propellant mass, and of the delaminations down to 1 mm between the thermal protection and the propellant. The total surface to be controlled for each segment is 100 m 2 . In order to speed up the examination, it is necessary to control a moving booster in front of the radiographic system. This real time control requires numerical radiographs and imaging methods to enhance the quality of the displayed radiographic images.
We have developed a system composed of a linear accelerator (15 MeV) and a 2D detector adapted to high energy X-rays (figure 1). The real time system called TOTEM allows to acquire radiographs during the rotation of the object, and to process them on line using a tomosynthesis algorithm. The system is installed at the industrial site of Kourou (French Guyana), and is currently being used by the operators in charge of the visual inspection of the boosters.
2.2 Tomosynthesis processThe principle of tomosynthesis is to accumulate the radiograph pixel values corresponding to the projection of the same point of the plane to be reconstructed. The capability to digitise radiographs allows to generalise this approach to non-parallel and non-linear geometries.
In the case of a cylindrical object moving around a rotation axis which is not in the projected area, the focal surface on which we can apply the process is the radial plane containing the rotation axis, which is orthogonal to the X-ray source/detector axis (Figure 1). For each point of the observed plane, its projection on the detector can be tracked during the object rotation, allowing to sum in the focal plane all the contributions of this point. The theoretical projection trajectory of each point of the surface on the detector is determined off-line by a calibration measurement. In this way we correlate the information contained in the projections [ANTO00]. The real time digital system TOTEM has been developed and tested. The dedicated system, controlled by a host PC, digitises, computes and displays a processed image at the video rate. All the images are digitised and the refresh ratio of the result image is 4 images per second in the tangential configuration. Figure 2 shows a typical radiograph of a tangential area of the cylindrical booster, and illustrates the improvement of the signal-to-noise ratio due to the tomosynthesis reconstruction.
3. The proposed method for defect extraction
3.1 Temporal considerationsThe system controls the booster by successive slices, each slice height depends on the size of the detector used (typically 40 cm). For each slice, the global sequence of radiographs corresponds to a complete rotation of the booster, during 6 min, thus approximately 10 4 images. A physical point of the object is visible – in the sense that it is projected within the detector – during a temporal interval that depends on the geometry of the booster and the rotation speed, typically 50 radiographs in our configuration. A typical delamination defect is located at the interface between the thermal liner and the propellant, and has the following shape : very thin in the radial axis (1 mm) and a tangential size that may reach several centimetres. Due to its shape and its contrast, such a defect is usually visible during 100 to 200 radiographs (figure 3). The method requires successive images to achieve robustness, but processing the whole sequence in a combined way is complex and time consuming. Thus a trade-off has to be found, defining the instantaneous past taken into account to status on the presence of a defect. This past length is directly linked to the duration of a defect visibility. The algorithm should ideally process on a sliding temporal mode. Of course the process has to be initialised properly.
3.2 Global schemeWhen processing a sequence, an image based method analyses each image, extracting what can be called alarms, and then combines those elementary results to conclude on presence of defects. By this way, the process focuses on the spatial level in a first step, then on the temporal level. This sequence processing approach has been recently used for X-ray inspection, with interesting results [MERY02]. For such techniques, the relative geometry of the different views has to be known precisely. Furthermore, the contrast of the defect should be sufficient to allow the detection of alarms on each single image.
In the opposite, other approaches in sequence processing apply temporal filters before considering the spatial aspect. This concerns the radioscopic mode, where a moving object intersects the X-ray flux. But generally, the successive radiographs can not be superimposed and compared at the pixel level. In the particular geometry of our configuration, the radiographs are comparable and a simple matching can be performed. In a perfect case, the interface appears at a static location within the sequence (generally vertical). If the rotation centre does not fit exactly with the center of the cylindrical booster – and this frequently happens due to the huge weight of the object, or if the geometrical axis of the object is not perfectly vertical, the resulting apparent displacement in the image plane is a horizontal translation. After correction of this displacement, the successive radiographs can be combined on large temporal intervals. We have defined a method combining both spatial and temporal aspects, performing successively :
By the way, the algorithm integrates the following prior information : a defect is a set of pixels, approximately spatially connected, slightly more brighter than the neighbourhood (lack of material), and visible on several successive images. We detail now the different steps of the method.
3.3 Temporal matching and differenceIn a first step, the proposed method computes the apparent local displacement between the current radiograph and a reference one. This reference image is acquired at the beginning of the rotation for the considered slice. Several radiographs are cumulated, in order to improve the signal/noise ratio in the reference image. It is supposed to be defect free (this can be eventually controlled by an operator). We use the same reference image for the whole rotation of the slice, because the potential displacement is around the nominal location, and the difference between the first image and the current one does not increase over time in the sequence. Other schemes have been considered, with a sliding reference image distant of an interval greater than the duration of a defect visibility. This alternative approach could be required depending on the configuration.
The apparent displacement is due to the non-perfect rotation positioning. If the shape of the interface is perfectly cylindrical, the displacement can be modelled by a uniform horizontal translation. Experimental tests with the boosters radiographs have shown that the displacement field is in fact not uniform due to the deformation of the insulation that covers the metallic wall. The displacement is computed by correlation techniques, taking benefit of all the contextual information :
The obtained displacement is applied to the current image to be analysed, and this registered image is subtracted to the reference one. Figure 4 illustrates this displacement. Notice the shape of a typical profile, where the interface between propellant and thermal liner is not obvious to locate precisely, due to the projection of a cylindrical object. But the algorithm uses only an approximate location of it, and computes the displacement accurately on the profile.
3.4 Instantaneous and cumulative confidenceThe following steps of the algorithm focus on a large strip around the approximate location of the interface (80 pixels). The purpose is to derive from the sequence a confidence measure at each pixel, quantifying the probability of the pixel to belong to a defect. First, an instantaneous measure is computed on each difference image (registered current image subtracted to the reference one) by a smoothed thresholding. As a matter of fact, this difference image contains noise, possible defects, and potential thin structures along the interface, due to a non-perfect alignment of the interface during the matching process. When considering the histogram of the image, we can notice that the main distribution of the grey values is due to the background. The contribution of the defect, if it exists, can hardly be distinguished as a second mode on the histogram except if the defect is large, but it generally disturbs slightly the histogram. The mean of the distribution is not necessary equal to zero, due to slow variations of the acquisition conditions during the rotation.
Consequently, we first estimate the parameters of the distribution on line : mean m from the maximum of the distribution, and standard deviation s from FWHM (full width at half maximum). Then we apply the smoothed thresholding (Figure 5) :
Finally, we get a measure at time t for a pixel which value is x : ct(x) between –1 and 1, where the value -1 (resp. 1) corresponds to a pixel surely in (resp. out of) the background distribution. A confidence map Ct is then updated by cumulating the instantaneous confidence (x) ct . For each pixel : Ct = sup(0,Ct-1+ct). By that way, the cumulative confidence is increased if the pixel belongs to the defect (instantaneous confidence close to 1) and decreased when it is outside the defect. Negative values are clipped to zero.
A typical defect will produce a confidence measure that increases during all the visibility of the defect, reaching its maximum at the end of the visibility period of the defect, and then decreases to zero. The maximum value of the confidence is linked to the number of images where the defect is visible. Notice that there is a temporal shift between the visibility period of a defect and the highest value of the confidence map.
3.5 Defect extractionWe are now provided with a temporal sliding confidence map where a high value corresponds to a pixel that has surely belonged to a defect during several successive images. It is to be hoped that a simple threshold could be applied to this map, but experiments have shown that a spatial aspect is required to achieve robustness : defects are composed of several pixels, connected by parts. Thus we proceed in two steps : first we apply a temporary labelling based on two thresholds : a pixel which value is upper than Th1 is labelled as defect, smaller than Th2 as background, as between Th1 and Th2 as doubtful. Th1 is linked to the number of images where a defect is visible, and depends strongly on the rotation speed and the geometry of the system. For our application 10 is a good value. The low threshold can be very low (2 or 3). In fact the method is not sensitive to these two thresholds thanks to the use of the third label doubtful. The second step consists in the application of a decision rule for each pixel, using a spatio-temporal neighbourhood : the decision for one pixel depends on the label of its surrounding pixels, but also on the decision of these pixels at the previous image [PIRE02]. At the end of the process, we get the connected components of the defect thanks to a connectivity analysis.
3.6 Defect characterisation in 3DOnce a defect is detected, it can be characterized in each image. We compute the following characteristics : convex hull, size and direction of each component. This allows to perform an additional simple sort, based on contextual information, eliminating for instance defects that are too small or not sufficiently aligned along the interface. This is done for each image. The defect, if it exists, being visible over several images, it is possible to characterize it in space and time in the sequence – this corresponds to a 3D characterization in the object frame. We perform this characterization by cumulating the binary maps of the defect all over the instants when it is detected, and compute shape parameters on this image. The temporal length of the defect, estimated from the maximum value of the confidence map averaged on the defect area, is linked to the tangential length of the defect along the interface.
4. Experimental results
4.1 Radiographs of boostersThe defect detection method has been tested on sequences of radiographs acquired at Kourou site on boosters under examination. For the moment, our algorithm is not implemented in a real-time version. Thus it should be tested off-line on sequences of radiographs previously stored. On the tested sequences, we have detected and characterized all the existing defects, and no false alarm. Nevertheless it is sometimes difficult to status on the accuracy of the parameters, as no one knows the true parameters of the defect, especially its thickness t which is very important. Figure 6 shows three examples : on top, a defect approximately 2 mm thick, the corresponding temporal length being 90 images (290 mm), partially disturbed by a permanent unsticking of the liner (but what is the actual limit of the defect ?), a second one on bottom left, very thin (1mm thick, temporal length 20 images, thus at the detectability limit), and the last one of thickness about 2 mm.
4.2 Simulated radiographsOn true boosters, defects are fortunately uncommon. Thus we have also used radiographs obtained by simulation thanks to the Sindbad Software developed in our laboratory [GUIL03], to establish the limits of the algorithm. The booster and the interface have been simulated, with possible deformation of the thermal liner. Various defects have been simulated, thickness down to the requirements limit of the control (1mm) and have been correctly detected (Figure 7).
4.3 Experiments on other cylindrical objectsWe have also tested our algorithm on sequences acquired from experiments on a X-ray low energy system, using a scale model composed of materials of scale density, and adding noise to be closer to the true high energy application (Figure 7).
5. DiscussionWe have presented a method combining spatial and temporal aspects to detect potential delamination defects with sufficient robustness in a sequence of radiographs. This method is required because of the poor contrast of the defect and the strong noise of the high energy radiographs, resulting in the fact that a single image is not sufficient to perform the detection. Our algorithm takes benefit of several prior information : cylindrical objects, possibility to detect the interface, location of the defect along the interface. Furthermore, the thresholds can be refined thanks to geometrical considerations as the rotation speed and the size of defect to be detected. Our algorithm may be simplified depending on the context, especially if the contrast of a defect allows to perform the detection on a single image.
It can be extended to other similar applications, typically tangential examination of rotating large cylindrical objects, especially for delamination detection. One part of the algorithm that should be adapted is the approximate location of the interface, which shape can be very different depending of the relative density of both materials. We have obtained promising preliminary results for the control of nuclear waste drums in concrete containers.
The authors thank M.Antonakios for his help, especially concerning the TOTEM system.