J. Jarmulak, Case-Based Reasoning for NDT Data Interpretation, PhD. Thesis, Delft University of Technology, The Netherlands, April 1999, ISBN 0-9535094-0-0. (Not in stock at amazon.co.uk but I still have paper copies available for the same price, or you can download it here for free (pdf))
Nondestructive testing (NDT) is a name for a range of methods and procedures used to determine fitness of industrial products for further use. Testing is necessary because such products may contain various defects either being the results of faults or inaccuracies in the manufacturing process or the result of use. Some of these defects can be easy to spot (certainly if they have already led to a failure) but some are not immediately visible. Their presence can be determined, without negatively affecting the inspected object, using special (NDT) techniques.
Many techniques, based on various physical phenomena, are used in NDT. Some of the most popular techniques make use of eddy currents, ultrasound, X-rays, gamma-rays, and acoustic emissions. An NDT system contains some energy source which emits energy that interacts with the inspected object. The interaction is registered using special sensors. The registered signal has then to be analysed to determine if it contains any indications of possible defects. The interpretation of data is often difficult and requires special training and expertise.
Because of increasing safety requirements, and because of the desire to extend the effective lifetime of products, an increasing amount of NDT inspection is being done. Ongoing mechanisation and automatisation of nondestructive testing procedures lead to an increasing amount of data acquired. This, in turn, creates a need for reliable automatic or automated data-interpretation techniques to reduce the workload on the operators. Though the use of automatic interpretation techniques serves, first of all, the increase in the speed of data interpretation, the increase in the reliability, precision, and reproducibility of the inspection are also important goals. Moreover, because computer systems are capable of analysing more data simultaneously than could be done by an operator, they can handle and interpret more complex data than an operator could.
In response to these developments in NDT, the ANDES (Automatic NonDestructive Evaluation System) project was started at TNO. The goal of the project was to investigate possibilities for the automation of NDT data interpretation and to demonstrate effective and feasible methods on two representative NDT problems. The emphasis was on investigating usefulness of Artificial Intelligence (AI) techniques. The two test cases which guided the research were: interpretation of data coming from an ultrasonic rail-inspection train (owned and operated by NS Ultrasoonbedrijf) and interpretation of eddy-current data from heat-exchanger inspection.
Artificial Intelligence provides a set of techniques for handling complex problems, usually needing human intelligence to solve. The most popular techniques are: artificial neural networks (ANNs) trying to model neural organisation of the brain, and rule-based expert systems which use expert knowledge represented in form of rules to solve problems. A more recent AI technique is case-based reasoning (CBR). CBR is based on the observation that people usually solve problems by reusing solutions that have worked in the past. This works because usually similar problems have similar solutions.
A CBR system stores previously solved problems, so-called cases, in a case-base. When a new problem has to be solved a search is done in the case-base and the most similar previously solved problem is retrieved. Based on the old solution, a solution to the new problem is determined. Because usually the retrieved problem is not exactly the same as the new problem, some adaptation of the old solution is necessary to fit the new problem. Once the solution has been verified to be correct, it can be saved in the case-base. By saving new cases a CBR system is improving its performance – it learns. Ability to learn is only one of the advantages of CBR systems. Other advantages are: relative ease of acquiring knowledge necessary to build the system (as it is contained mainly in cases), high reliability due to use of cases as contextualised knowledge representations and the capability to recognize new situations, reduced need for maintenance as the system can adapt to many changes in the problem domain. If maintenance is necessary it is relatively easy, usually confined to the maintenance of the case-base.
The advantages of CBR seem to make it especially suited for NDT problems, where the knowledge about data interpretation is often difficult to formalise, the reliability is crucial, and ease of maintenance is important from the economic point of view. Because CBR systems learn they can adapt to changing inspection conditions. Moreover, the normal way of doing data interpretation by NDT inspectors is, to a large degree, CBR-like.
Within the ANDES project two prototype CBR applications have been developed, one for each of the already mentioned test cases. The tests conducted on data from real field measurements were satisfactory for both systems. The CBR system for URS is currently being prepared for use on the inspection train.
During the development of the two systems, the issues of optimal case representation and case matching demanded much attention, as they are crucial to the retrieval of the most similar cases of the already classified data from the case-base. Related to this is appropriate data preprocessing, like removing noise and other unwanted influences, which makes the data easier to interpret. Also, obtaining data necessary for testing of the system proved to be far from trivial. Moreover, the URS system demanded considerable effort in optimally designing case-base to achieve acceptable retrieval times.
Though the two developed prototypes, as well as other systems described in the NDT literature, demonstrate the capabilities of AI for NDT data interpretation, the use of such systems for real inspections is still almost nonexistent. One reason for this may be the complexity of the factors involved in the decision making about NDT inspection. Though, in principle, it is possible to demonstrate the advantages of using automated data interpretation systems (or even fully automated inspection systems) in clear economical terms, such analyses are complex, and make sense only for particular test cases (generally applicable analyses are not feasible). Moreover, introducing new techniques (not only for data interpretation) in NDT is made difficult by a common involvement of at least three parties in the inspection: the inspection-system manufacturer, the inspection company doing the actual inspections, and the owner of the inspected product or installation. Successful introduction of new systems requires cooperation of all the parties and it is not easy to achieve. Possibly, if automated data interpretation systems become cheaper and easier to construct and to use, without compromising reliability, then their introduction becomes easier. CBR is a technique which may help to achieve this.