You are on page 1of 8

7

Perspectives of NDT Data Fusion

The use of diverse methods, in concert, to solve data fusion problems is still evolving
DL. Hall, 1992

7.1 Concluding comments


The most common data fusion methodologies, theoretical and experimental approaches to
the concept of NDT data fusion, are presented in this book. The identification of data
fusion algorithms and of the factors that contribute to the reliability of NDT data fusion
are discussed. The design of a data fusion system for industrial applications is described.
This chapter completes the work presented on multisensor data fusion by summarising the
material from earlier chapters and making concluding remarks on the theory and
application of NDT data fusion. Future research directions and possible applications are
also presented for the further development of multisensor NDT data fusion.
Application of NDT data fusion, described in chapters 5 and 6, showed that fusion
should be performed using data which is as close as possible to the original data. This has
the effect of minimising loss of information, which could occur from extensive signal
processing, and reducing complex processing operations. The criteria for the design of a
data fusion model are identified as maximum simplicity, maximum efficiency and
operational flexibility. Unfortunately the last point is not a synonym for simplicity. Raw
sensor data constitutes the most common input format of data integration and fusion
systems; however, if a multi-usage compatible system is required, other parameters have
to be considered. Among these are information on sensor location, information on external
databases and environmental data, as well as information related to the experience of
human operators (Fig. 7.1). Therefore very complex data management would be required
to develop such a powerful system to perform the following operations:
store data in multiple format;
access and modify data format if required;
allow interactive access of data by multiple users;
have a user friendly interface;
be compatible with other systems;
be secure to prevent unauthorised access to the data;
be able to compress data for optimisation of storage space.
It is clear from the research presented in this book that multisensor NDT data fusion is
Defect Detection

Defect Quantification NDT Techniques NDT Equipment Component Inspected Inspection Location Inspection Procedure

Location Orientation Length Depth Visual MPI .etc. Specification Calibration Material Dimensions Access On site Laboratory Manual Automated

Environment Inspector

Qualification Experience Age

Fig. 7.1 Factors to consider in the development of a data fusion system for multiple uses
worthwhile and can improve the overall efficiency of a non-destructive examination. It has
also been demonstrated that NDT data fusion could be applied to different types of
inspections, regardless of sensor types. The data fusion algorithm implemented took the
form of probabilistic inference processes such as the Bayesian inference theory, the
Baysian estimate and Dempster-Shafer evidential reasoning. Data fusion can be performed
at the signal level using raw eddy current data collected during the inspection of composite
materials and with data which characterise defect size (depth or length) from eddy current
and ultrasonic sensors used for weld inspection. A pixel level data fusion approach with
signals from eddy current and ultrasonic systems applied to weld inspection has also been
used. It was demonstrated that data fusion can help to make an inference about a hypothesis
both in the case of binary decision making, such as for defect/no defect1, and in the case of
quantitative information, such as defect depth, from more than one NDT system.2 From the
experimental results achieved, the Dempster-Shafer approach was preferred as it is more
efficient than the Bayesian approach in making accurate estimations of defect depth. It also
presented the results with associated probability intervals which were used to make
decisions in favour of a data fusion output with the highest degree of confidence. The
outcome of the Bayesian inference process is a posterior probability which either supports
or refutes a hypothesis. This type of reasoning is useful for binary testing and provides a
measure of uncertainty of a hypothesis. In the case of non-binary testing, the
Dempster-Shafer evidential reasoning is better suited to making inferences. In both cases,
an increase in sensor number would not only reduce the inspection time but also increase
the performance of the fusion system. It was also noted in chapter 6 that the final decision
outcome is affected by the performance of each NDT system. Improvements from data
fusion would only be achieved if each NDT system is in a similar performance range. No
advantage would be gained from the fusion of NDT systems with poor performances as the
belief associated with any sensor output would be very low.
As stated by J.W. Tukey (1977), 'We have not looked at our results until we have
displayed them effectively'. Visualisation of NDT data was performed using commercially
available software which provides an affordable way to display data in a format which has
colour coded images. This allows rapid location and sizing of defects as illustrated in
chapter 5, with the detection and visualisation of disbonds and impact damage in
composite materials. For NDT data fusion, visualisation enabled clarification of many
aspects of defect detection by presenting data fusion results in a format which facilitates
interpretation and also enables pixel level data fusion. Similar displays could be used as
inputs to neural networks to assist the human operator in decision making by performing
pattern recognition tasks as well as pixel level data fusion operations.
Prior to the implementation of NDT data fusion, the most common NDT methods were
described and their physical principles studied before carrying out experimental
inspections. This brief study is necessary to select complementary NDT methods, the data
from which could be combined at a later stage. The use of eddy currents to detect and
quantify defects in composite materials was also investigated. This study demonstrated that
electromagnetic techniques were a low cost, highly reproducible and efficient alternative to
ultrasound and infrared thermography for inspection of composites. Standard statistical
analyses such as POD and ROC were used to assess the performance of the eddy current
system. Similar statistical analyses were carried out to assess the performance of the data
fusion algorithms implemented, which demonstrated - in the experiments described in
chapter 6 - the Dempster-Shafer evidential theory to be more accurate regarding decision
making of defect depth estimation than the Bayesian theory.
From the comment by R.A. Armistead, 4No single NDT method alone can provide a
total solution to the needs of the advanced engineering materials community', the use of
multiple NDT techniques is anticipated, leading to a need to display and combine data
effectively. Fusion of NDT data can be used to combine information from multiple
identical or different sensors and to make inferences on inspection results. The technology
is already available to perform such a task but only a direct requirement by industry would
promote further development of NDT data fusion. Research has already started at an
industrial level and it is not surprising that this need has come from the nuclear industry
for which the use of multiple NDT methods is necessary to meet safety standards.
More research is required to develop an NDT data fusion system for on-site inspections,
to develop a database and a man-machine interface and to configure a system for a
specific application. Configuration and interfacing of communication software for data
collection and data transfer, and of visualisation software for data display, mapping and
analysis would be required. A data fusion algorithm adapted to the problem would have to
be selected or especially designed, depending on the data format and application.
The statistical approach to NDT data fusion described in this book demonstrated that an
improvement in defect characterisation can be achieved by combining information from
multiple sensors. Two approaches, based on the Bayesian and the Dempster-Shafer
theories, were implemented for NDT data fusion. From these two approaches, it was noted
that the Bayesian posterior probability tended to be more affirmative than the support
generated by Dempster-Shafer, because in the latter, when there is conflicting informa-
tion, uncertain results are obtained. The Bayesian inference process provides a probability
of a hypothesis being true when given evidence; however, the prior probabilities are
highly dependent on experimental results. Unlike the Dempster-Shafer approach, there is
no uncertainty associated with the decision towards a hypothesis resulting from a Bayesian
calculation. However, the Bayesian approach is more appropriate for binary testing with
multiple identical sensors than the Dempster-Shafer theory. Owing to the difficulty in
defining prior probabilities, the Bayes estimate approach appeared to be best suited to
making inferences towards measurement of an unknown quantity, i.e. a defect depth or
length. It was noted that the Bayes estimate has the advantage of producing an estimated
defect depth when given evidence from multiple sensors of which a normal prior
probability on n and a can be assumed. The Dempster-Shafer process is highly dependent
upon the sensor efficiency; a small change in input data can produce a large variation in the
outcome of the Dempster-Shafer rule of combination. For example a 55 per cent increase
in accuracy for length estimation can be achieved with the Dempster-Shafer rule of
combination when combining length information from instruments in relative agreement.
But a variation of 1.00 mm of length measurement from one instrument can produce a
change in the output, and an increase of only 11 per cent in accuracy is achieved.
Combining similar information using the Bayesian approach produces an increase in
accuracy of 38 per cent in both cases, regardless of the sensor variations. The advantages
and limitations of both approaches are summarised in Tables 7.1 and 7.2.
The data fusion process is dependent on the type of defect detected and the equipment
used. Data input is in the form of length or depth measurements, and preprocessing of the
original signal into a common numerical format is required to build a data fusion engine
which will be able to combine information at the signal level. A large amount of
experimental data needs to be collected prior to fusion, on multiple test samples and with
several NDT instruments, in order to be able to build a database to assign prior
probabilities to each measurement. However, once this set of data has been collected, data
fusion could be performed in real time using a computer program specifically designed for
a particular inspection procedure. A schematic for the development of an NDT data fusion
engine is presented in Fig. 7.2.
Table 7.1 Advantages and limitations of the Dempster -Shafer data fusion approach

Advantages Limitations

Provides a soft decision output Poor results if both systems are in relative
Associates belief and uncertainty values to disagreement
a decision output Small changes in input data can produce important
Better accuracy than Bayesian approach changes in decision output
for depth estimation (5% more accurate) The estimated defect depth is given as a depth interval,
not a definitive depth value

Table 7.2 Advantages and limitations of the Bayesian data fusion approach

Advantages Lin:iitations
Best suited for binary decision testing Provides only a hard decision output (no uncertainty values)
Low computational requirements Prior probabilities can be difficult to obtain (Bayes estimate
is preferable)
Output highly dependent on sensor efficiency
Sensors need to be of similar efficiency (not suitable for
combining information from sensors with high discrepancy)

7.2 The future of NDT data fusion


From the progress achieved in the fields of NDT data visualisation, artificial intelligence,
pattern recognition, NDT data fusion and discussions with industries in need of an efficient
and reliable NDT technique, a rapid evolution of the concept of NDT data fusion is
foreseen.3'5 It is more economically viable to keep using conventional NDT equipment
and combine data collected from such systems than to redesign existing apparatus. One
single NDT instrument can be used for multiple inspection purposes and data produced
combined using a data fusion system designed for a specific NDT task. Pattern recognition
and signal processing are already used to build expert systems to detect and classify
defects with minimum operator intervention. In a highly advanced NDT data fusion
system, the expertise of the human operator could be coupled with information provided
by a machine, and expert systems developed or adapted to represent this knowledge and
make inferences. The development of a neural network could have four major advantages:
• to estimate the certitude of the sensor output (NN input) for different NDT
systems;
• to extract significant information from each system;
• to extract relevant information from each system in relation to the type of
inspection;
• to indicate a decision in a numerical or graphical format easy to interpret by an
operator and ready for fusion.
Identify Defect Location

Defect
Examine Geographical Type
Sensor Location Elements

Defect Extract Data Database


Sensor 1 Signal
Detection
Information Alignment Management

for Data Fusion Decision

Prior Cutout
Engine
Probability
Defect Extract Data
Sensor N Signal Information
Detection
Information Alignment

Identify Sensor Type Defect


Type
Elements

Identify Defect Location

Fig. 7.2 Schematic for development of a NDT data fusion engine


A three-level expert system could be designed to assist in non-destructive examinations
(Fig. 7.3).

Level 1 Level 2 Level 3


NDT Sensor 1
Data
Alignment Data Fusion Situation
Assessment
Centre
Signal Failure
NDT Sensor n Processing
Analysis

Human-Machine Interface

Decision
Fig. 7.3 Design of an expert system to assist in non-destructive examinations

Signal processing, data alignment and correlation are performed at the first level while
data fusion operations are carried out at level 2 using conventional or specific data fusion
algorithms such as Bayesian, Dempster-Shafer theories or fuzzy logic. At level 3, an
expert system is used for situation assessment, identification of defect type, estimation of
defect size, location and orientation. Other tasks such as failure analysis can be performed
and information collected from an inspection compared to that of pre-existing information
stored in a database. A human-machine interface analyses equipment malfunction and
displays signals from NDT systems in an analogue, digital, 2-D or 3-D colour coded
format, as well as displaying information from the data fusion centre in the form of images
and/or statistical and probabilistic numerical values. The operator is also informed of the
output of the expert system operations and of any major danger which may be associated
with the presence of a defect. The final decision is left to the human operator but could be
automated if required.
Fully automated inspection using a remotely operated vehicle (ROV) for testing in
hazardous environments has already been developed for the nuclear and offshore
industries. By fitting multiple NDT sensors to an ROV, automatic logging of data, display
of remote information in a safe environment and real-time data fusion could be performed
(Fig. 7.4).
Instrument and operator errors are significant factors in sizing and location of flaws. The
knowledge and expertise of a human operator could be coupled with information provided
by a machine in order to reduce human error. Artificial systems could be developed further
to represent this information and to make inferences. Nuclear plants already have
supervisory control configurations where decision support systems using Dempster-Shafer
theory aid human operators to perform high level tasks by processing and displaying
uncertain information.6 The final phase of this approach could be the adaptation of multiple
NDT systems on remotely operated vehicles for completely automated inspection.
Safe Environment

Data Logging Data Visualisation Data Fusion

Ultrasonic
Sensor

ROV
Eddy Current
Sensor

Underwater Environment
Fig. 7.4 Schematic diagram of a fully automated ROV inspection

References
1. Gros XE, Strachan P, Lowden DW. A Bayesian approach to NDT data fusion, May
1995, Insight, 37(5), 363-7.
2. Gros XE, Strachan P, Lowden D. Theory and implementation of NDT data fusion,
1995, Research in Non Destructive Evaluation, 6(4), 227-36.
3. McNab A, Dunlop I. A review of artificial intelligence applied to ultrasonic defect
evaluation, Jan. 1995, Insight, 37(1), 11-16.
4. Kirk I, Lewcock A. Neural networks - an introduction, Jan. 1995, Insight, 37(1),
17-24.
5. Windsor CG. Can we train a computer to be a skilled inspector?, Jan. 1995, Insight,
37(1), 36-49.
6. Hasegawa S, Inagaki T. Dempster-Shafer theoretic design of a decision support
system for a large complex system, 1994, Proceedings of Institute of Electrical and
Electronics Engineers International Workshop on Robot and Human Communication.

You might also like