Abstract

In the frame of the Plant Life Assessment Network (PLAN European Thematic network), a working group has been organised around the theme of «Data Analysis and Fusion». This was a mean to develop scientific exchanges between European projects dealing with similar problems. In this paper, the interest of data fusion is shown through the results of two Brite projects, where one area of innovation was to combine automatically the data coming from two NDT methods (X-ray and ultrasonic techniques). The interest is to deliver a global inspection report on the sample under test, and thus increase reliability. These projects ended respectively in February 2000 and December 1999, and the data fusion was validated.

Introduction

In more and more inspection applications, we need more than one type of measurement to tell us everything we want to know about a component, a process, ... Moreover, even one measurement must be as reliable as possible, and reliability must be quantified. To this end, data fusion is a powerful tool to combine data in a reproducible and objective way. A lot of problems (inspection, material characterisation …) can not be solved on the basis of only one set of data, but require the use of different measurements. The «data fusion» refers to the scientific combination of the measurements. In some cases, different data sets (measured by different types of sensors) give information on various aspects of the problem : they are said to be complementary. The interest is thus obviously to obtain more information than by using only one measurement. On another hand, the information provided by the various data sets can also be redundant, e.g. dealing with the same aspect of the problem. This could be considered useless at first glance, but actually this improves reliability, because the information given by one measurement is confirmed by the other. Complementarity and redundancy of data sets are the basis of a Data Fusion application. It must be noticed that data fusion implies n sensors of different types, by definition. In the PLAN network a wide application range of data fusion is found among the projects:

• localisation and characterisation of defects inside a component,
• corrosion monitoring,
• comparison of inspection result and CAD description of a part
• linking process monitoring to finish product quality,
• condition monitoring of rotating machinery
• characterisation of a dynamic process with several components
• estimate of crack growth

It can be easily understood that those applications are very different from each other, this is why in principle, each problem has to be solved with a specific way of combining data (neural nets, mathematical average …). Sometimes, a priori information about the process or the component to be characterised is required.

In this paper, one particular type of data fusion is detailed, in the field of weld inspection, using radiographic and ultrasonic tests. The same theory (Dempster-Shafer) for combining data was used in two Brite projects that ended successfully.

Use of data fusion in weld inspection

The inspection of welds often requires both radiographic and ultrasonic testing, in order to increase the reliability of inspection. In both the FFRESHEX [Fast Film REplacement System for High Resolution X-ray weld inspection with ultrasonic data fusion, BRITE/EURAM BE 3681] and the MISTRAL [Multi sensor Inspection System for component Testing : toward more Reliable non destructive testing AppLications, BRITE/EURAM BE96 - 3482] projects one area of innovation was to combine automatically the data coming from both NDT methods, in order to provide a complete inspection report on the sample under test, and thus increase reliability. These projects ended respectively in February 2000 and December 1999, and the data fusion was validated.

A fusion approach is an intelligent combination of the information given by the various testing methods. Several theories of data fusion exist, among which, in the field of NDT, the evidence theory is known to be adequate, because of its ability to model both inaccuracy and uncertainty (1-5). Another interest of this theory for NDT applications is that it is possible to discriminate between two hypotheses. In our case the hypotheses are simply «there is a defect» and «there is no defect». The discrimination between both hypotheses can be quantified in terms of the reliability of the hypothesis arrived on whether the inspected area is sound or not. The doubt between two hypotheses is also modelled in this theory, through the combination of hypotheses (in our case «defect or no defect»). This possibility to model combined hypotheses is an originality of the evidence theory.

Registration of data sets in a reference system

The choice of a common data format and coordinate system is the first and sometimes the major problem to solve in the development of a data fusion task. Specific solutions have to be found depending on the kind of data concerned, and there is no generic tool for that. In the field of NDT, a data format was created in the TRAPPIST project which was then used in MISTRAL. In the FFRESHEX project the registration was facilitated by the fact that a combined acquisition system was developed in the frame of the project.

Data fusion in the FFRESHEX project

In the FFRESHEX project, the information available concerns ultrasonic signals (B-scan images from several probes) and X-ray image(s) of the same volume of the sample.

Ultrasonic acquisition and processing

8 probes are used in pulse-echo mode in order to gather information of the whole weld volume. Each probe gives an image such as that shown in the figure 1.

Fig 1: Ultrasonic image given by one of the probes. The zone corresponding to the weld is between the two dark lines. This is a B-scan image, that is, the horizontal direction represents the abscissa along the weld, and the vertical is the time of flight.

The signals from the 8 probes are analysed and processed. Echoes coming from the same position inside the weld can sometimes be detected by several probes. In this case, they are gathered in a single object (probes merging). The amplitude of each echo is compared to that of a reference echo from a known hole and thresholds were defined in order to classify the echoes with respect to the reference. The result of the ultrasonic stage before combination is thus a list of detected indications with their respective amplitudes and positions inside the weld volume. The position is not a single point but a volume around each position, depending on the object size and uncertainties related to the measurement.

X-ray acquisition and processing

A new detector was developed in the project allowing the acquisition of high resolution images. Figure 2 displays a typical radioscopic image of a weld, with a pixel size of 54 µm.

Fig 2: Part of a weld with an artificial slot of 0.5mm in width and 25mm in length.

An automatic image processing method was developed to detect the defects, based on adaptive thresholding. When developing a segmentation algorithm, one wishes to be able to discriminate (by image processing) as many defects as possible with the least possible number of false indications. Weld images are particularly difficult to process due to the weld cap, which is non-uniform, and irregularities of the weld. These irregularities can be confused with defects, and thus, yield some false indications. Thus, if the image segmentation algorithm is very sensitive, a high uncertainty is connected with the assessment of a defect, and the false alarms rate increases. In our study, we take advantage of the data fusion to sort false alarms and real defects, and thus, a sensitive image processing was preferred. The segmentation algorithm is detailed in (6). The sensitivity of the method allows to detect porosities down to 100 µm in diameter, which is quite good as compared to the pixel size (50 µm). Once the objects are detected, some features are computed for each object, namely area, contrast to noise ratio, position and elongation. These features will be used in the data combination stage.

Data combination (fusion)

The information obtained from each non-destructive testing (NDT), inspection or characterisation method is a list of objects that have been detected, and their associated parameters (amplitude, contrast…). From these parameters, it is necessary to compute a confidence level associated to each NDT method. This is called a mass function in the Dempster-Shafer or evidence theory. This confidence level can be compared to the confidence of the radiographer when he looks at a film and says “I am more or less sure that there is a defect”. In the data fusion algorithm, the subjective “more or less” is quantified to a precise value. The process connected to this decision-taking is crucial to the data fusion algorithm. In our study, for each testing method, the computation of the confidence levels is done after an apprenticeship stage where the knowledge of the NDT experts is used (7).

The fusion stage starts by an object matching procedure in order to check whether one indication has been detected by both techniques. Then, the masses (or confidence levels) issued from each method are combined by the orthogonal sum of Dempster, which formula is illustrated on the figure 3. The masses before fusion are plotted in a two dimensional graph. The cross products between the masses correspond to areas of the graph, and allow to compute the masses after fusion. A conflict occurs in case the two methods give a mass on two hypotheses having no intersection because in this case the methods are in contradiction. This can help in the follow-up of the good working of the techniques : if the methods are in contradiction, the value of conflict will give a warning to the operator.

On another hand, the doubt of one method does not deny the confidence of the other method, that is why the confidence after fusion can be higher than none of the method alone.

Fig 3: Illustration of the data combination by the orthogonal sum of Dempster.

The figure 4 illustrates the entire process on a porosity. In this example, we see how the redundancy of the methods can improve reliability of inspection : the two techniques have detected the porosity, but with a confidence of respectively 70 % (X-ray) and 50 % (US). After fusion, the confidence is 85 %.

Fig 4: Illustration of the data fusion process for a porosity.

Results

The system was tested in a spool yard in Norway in November 99. The feasibility of the combined acquisition system was proved. The following part explains the results of the data fusion process on a pipe with a circumferential weld. The plate material is carbon steel and the plate thickness is 14.5 mm. It contains one external slot, two internal ones, one small area with lack of fusion, some undercuts and a few number of porosities. This pipe has been repaired several times so that the weld cap is very irregular, which means that it corresponds to one of the most difficult inspection case on site. However, this is an interesting case for performance demonstration, because the weld contains both artificial and real defects.

After image processing, all the objects detected on the radioscopic image were compared to the defects identified by an expert on the radiographic film (in order to assess the quality of both the radioscopic image and the image processing). This film was made at the spool yard with a radioactive source located inside the pipe (panoramic single wall image with a Selenium 75 source). In the following figures, all the true defects are surrounded by an ellipse (true means that the objects correspond to a defect detected by the expert on the film), and the others are said false. The X-ray segmented image (Fig. 5) contains 6 false indications corresponding to high local variations of the weld thickness. Most of the porosities are detected and the three artificial slots are also detected but the lack of fusion cannot be detected because it is very small and has a very low contrast. We can also distinguish some undercuts in the lower right part of the image detected as many small objects. The subsequent images represent the defects detected by both methods, where the image width is the external circumference of the pipe and the two horizontal lines define the external weld cap limits. Both images are in the same reference system for the defect position and each square of the grid is 5 mm x 5 mm.

Fig 5: Objects detected by X-ray inspection (after image processing and registration in a common reference system).

Concerning the defects detected by ultrasonics, all slots are detected and also 3 undercuts and one lack of fusion (Fig. 6). None of the porosities can be detected due to their very small size. This image contains only one false indication. We must say here that the object is said false because it is not visible on the X-ray film, because it is the only reference we have. But it might also be a lack of fusion, missed on the X-ray film.

Fig 6: Objects detected by ultrasonic inspection (after processing, probe merging and registration in a common reference system).

The complementarity of the methods appears here, as it was impossible to detect all the defects with only one method. Thus, the multi-technique approach gives more information about the inspection. However, in terms of data fusion, only the defects that are detected by both methods can be fused. After the fusion stage, the results are displayed for the operator, as in figure 7. The final decision is left to him, but the confidence level is given by a colour: red for a high confidence (from 67 % to 100 %), orange for a medium confidence (from 33 % to 67 %) and green for a very low confidence (below 33 %). The image of the weld which was displayed in a linear format in figure 5 and 6 is in fact a circumferential weld.

Fig 7: Operator display interface: top left window : section view of the weld volume; top right window : view of the pipe circumference; bottom window: weld top view. Two cursors allow selecting a part of the weld to be displayed. A frame indicates the decision on the selected defect as a colour (red for surely a defect, orange for probably a defect and green for surely not a defect). The lowest picture is a zoom over a part of the circumference, after defect matching. The section view of the pipe include both figures 5 and 6 on the same co-ordinate system.

After the registration in a common reference co-ordinate system, the objects are then displayed in both a section view of the pipe (top right diagram), and a linear view (lower diagram). This view is a zoom of a part of the circumference which is defined by the operator. Data fusion allows for detection of more defects than with any of the methods alone, and the display of the confidence associated to the defect is a great help for the operator, as well as the display of the defect location inside the weld volume.

Data fusion in MISTRAL

The Fusion module of the MISTRAL application is based on the same fusion architecture, a high-level architecture, where the single processings from each technique are performed first, and on the same mathematical model, the Dempster-Shafer theory, as FFRESHEX. Some particularities and differences can nonetheless be listed. Firstly, one difference in the MISTRAL project with respect to FFRESHEX concerns the data. Several X-ray images are used, in order to have a three dimensional information. The 3D information obtained via these X-ray images is not complete, though, as there are only small angles between the various positions of the X-ray sources. This constraint is introduced to simulate an on-site situation, where the inspected component may not be easily accessible. Second, the method developed is also based on the description of uncertainty via the Dempster-Shafer model. The difference with the Ffreshex algorithm is that in Mistral data fusion is done on a pixel level, whereas in Ffreshex it is on an object level.

In order to enable the data fusion, a common description has to be found for the different kinds of data. We are working in a three-dimensional framework: this common description has to apply on each voxel of a reconstruction volume. The first step of our work is therefore to produce what we call “evidence values” from both the ultrasound and X-ray data, for each considered voxel. These evidence values account for the certainty, given by the data, of presence (positive evidence) or absence (negative evidence) of a defect in the inspected component. The uncertainty is measured by the “doubt” evidence value. For each kind of data, we therefore obtain a set of three evidence values (positive, negative and doubt) in each voxel of the reconstructed volume. The calculation of these evidences, for one inspection technique, is based on a comparison between the global distribution of the amplitudes (on the whole data set) and the local distribution (the concept of « local » relying of the definition of a neighbourhood, different for each technique). We then apply the orthogonal sum of Dempster and Shafer to these two sets of evidences, combining them to produce a final set of evidence. For each voxel of the reconstruction volume, we therefore obtain a set of evidence values that take into account both the ultrasound and the X-ray data. Another difference that has to be noted between the Mistral and the Ffreshex approach is the handling of the conflict between the sources. Ffreshex considers the factor K, representative of the conflict, as a warning to the operator, whereas in Mistral, the algorithm has been adapted to take the conflict into account directly in the evidence values. Let us recall the calculation of the various factors following the orthogonal sum of Dempster-Shafer, as stated above, figure 3. In our example, we have:

where P_x, N_x and D_x are respectively the positive, negative and doubt evidence values related to the source « x », either X-ray data (XR) or ultrasound data (US).

The idea is the following. The factor K, in the Dempster-Shafer model, measures the contradiction between the various sources. The more K is near to the value 1, the greater contradiction there is between the various sources. When the evidence masses related to each source are very dissimilar (K very near to 1), we may have to reconsider the result of the combination. Therefore we introduce a “modified combination rule”:

This new formula is interesting when there is a great dissimilarity between the results of the various sources (when the factor K is near 1), because in this case it increases the doubt, which corresponds to the limited conclusions we can draw from the various sources. Tests have been performed on data acquired on mock-ups, one representative of the problematic of the energy industry, and a second mock-up, a welded pipe, representing one of the monitoring problems of the chemical industry. The obtained results show that it is possible to take advantage of the complementarity of the two kinds of data. The localisation and characterisation of the defect are improved by combining the advantages of ultrasound data (good depth resolution) and of X-ray data (good lateral resolution).

Conclusion

By detailing one practical example (weld inspection), our aim was to put the emphasis on how reliability of inspection can be enhanced thanks to data fusion. The interest of data fusion relies first on the use of complementary methods: more information about the sample is thus available than with one method only. Another interest of data fusion is the improvement of reliability by using the redundancy of the methods. This happens when several methods detect the same object with a rather low confidence, then, of course the data fusion presents a great advantage because the confidence after fusion increases. In other words, data fusion really improves reliability of defect detection when there is a part of redundancy between the methods, whereas the complementarity of the methods adds information but does not influence the reliability. Moreover, with this type of Data Fusion Architecture, the reliability of the single processes has to be taken into account when evaluating the fusion result. The work done in this Cluster task of the Plant Life Assessment Network was successful in terms of sharing knowledge between the members, and it was especially interesting for projects dealing with similar techniques like data fusion. In particular, the various solutions to some problems occuring frequently when performing data fusion : geometry handling, alignment, data format, synchronisation in space and time,... have been shared.

Acknowledgements

We would like to thank all the partners of task A1, especially Brian S. Hoyle and Sophie Goujon, for the useful discussions we had, and also the PLAN co-ordinator and secretary, Roger Hurst and Darren Mc Garry, for their interest in our work.

References

1. X.E. Gros, NDT Data fusion, John Wiley and Sons, New York, 1997, 205p.
2. D. L. Hall, Mathematical Techniques in Multisensor Data Fusion, MA: Artech House,1992,.
3. T. Lee, J.A. Richards, P.H. Swain, Probabilistic and evidential approaches for multisource data analysis, IEEE trans. on Geoscience. And Remote Sensing, vol.25, 1987, pp. 283-293.
4. R.C. Luo, M.G. Kay, Multisensor integration and fusion in intelligent systems. IEEE trans .on systems, man, and cybernetics, Vol. 19, no 5, 1989, pp. 901-931.
5. G. Shafer, A mathematical theory of evidence, Princeton University Press, Princeton, New Jersey, 1976.
6. O. Dupuis, Fusion entre les données ultrasonores et les images de radioscopie à haute résolution : application au contrôle de cordon de soudure, PhD INSA de Lyon, 2000.
7. O. Dupuis, V. Kaftandjian, D. Babot, Y.M Zhu, Automatic detection and characterisation of weld defects : determination of confidence levels for data fusion of radioscopic and ultrasonic images, Insight : Non Destructive testing and Condition Monitoring , vol. 41, n°3, 1999, pp.170-172.

© NDT.net - info@ndt.net

|Top|