ENIQ RECOMMENDED PRACTICE

ENIQ Recommended Practice 13

Qualification of Non-Destructive Testing Systems that Make Use of Machine

Learning
Issue 1

ENIQ Report No. 65

Technical Area 8
European Network for Inspection & Qualification

June 2021
SNETP Association
c/o EDF
Avenue des Arts 53B, 1000 Brussels, Belgium
Email: secretariat@snetp.eu
Website: www.snetp.eu

This document was prepared by the SNETP Association.

LEGAL NOTICE
Neither SNETP nor any person acting on behalf of SNETP is responsible for the use which might be
made of this publication.
Additional information on SNETP is available on the Internet. It can be accessed through the SNETP
website (www.snetp.eu).
Brussels: The SNETP Association

ISBN 978-2-919313-28-0
 ENIQ RP13 – Qualification of NDT Systems that Make Use of ML

Index

Executive Summary ................................................................................................................................. II

1. Introduction..................................................................................................................................... 1
2. Objectives ........................................................................................................................................ 1
3. Machine Learning as Part of a NDT System .................................................................................... 1
3.1. Artificial Intelligence / Machine Learning Fundamentals............................................................. 1
3.2. Cases of Use and Extent of Qualification ..................................................................................... 3
3.3. Challenge and Opportunity of Machine Learning ........................................................................ 4
4. Qualifying Machine Learning NDT Systems..................................................................................... 5
4.1. Example of ML Qualification Flow Chart ...................................................................................... 5
4.2. Demonstrating the Performance of a ML System ........................................................................ 5
4.2.1. Specific Aspects for the Technical Justification ................................................................. 5
4.2.2. Qualification Practical Trials .............................................................................................. 7
5. Summary ......................................................................................................................................... 8
References ............................................................................................................................................... 9
Appendix: Machine Learning Fundamentals ......................................................................................... 11
Glossary ................................................................................................................................................. 18

ENIQ – European Network for Inspection & Qualification I

 ENIQ RP13 – Qualification of NDT Systems that Make Use of ML

Executive Summary
This Recommended Practice (RP) has been developed as a consensus document amongst the members
of NUGENIA Technical Area 8 (TA8) – European Network for Inspection and Qualification (ENIQ). The
main objective of this RP is to support licensees, qualification bodies and inspection vendors to
produce and assess inspection procedures that use machine learning (ML) for automated data analysis.
For the most part, the qualification of non-destructive testing (NDT) systems that utilize ML is similar
to qualifying more traditional NDT systems. This document provides guidance on the specific
considerations related to the use of ML in the qualification process.

ENIQ – European Network for Inspection & Qualification II

 ENIQ RP13 – Qualification of NDT Systems that Make Use of ML

1. Introduction
The European methodology for the qualification of non-destructive testing [1] is intended to provide
a general framework for the qualification of inspections of specific components to ensure that they
are performed in a coherent and consistent way while still allowing qualifications to be tailored in
detail to meet different national requirements.
This ENIQ Recommended Practice (RP) will assist those involved in the qualification of non-destructive
testing (NDT) systems that use machine learning (ML) models as part of the data analysis procedure.
This RP is relevant to any inspection method and builds on the previously published ENIQ position
paper “Qualification of an Artificial Intelligence / Machine Learning Non-Destructive Testing System”
[2].

2. Objectives
The main objectives of this RP are to:
• Identify the specific challenges related to the use of ML data analysis as part of qualified NDT
systems;
• Show the differences to a qualification of a conventional NDT system;
• Promote the harmonisation of practices in qualifying such systems; and
• Provide guidance on how to address the specific challenges and how to qualify such systems
and procedures.
Although this document is developed specifically for in-service inspection (ISI) of nuclear power plant
(NPP) components, the principles given here can also be applied to the qualification of manufacturing
inspections or ISI performed for non-nuclear applications.

3. Machine Learning as Part of a NDT System

3.1. Artificial Intelligence / Machine Learning Fundamentals
Artificial intelligence (AI) is the field that aims to develop systems that mimic human intelligence or
perform tasks that have been thus far thought to require human intelligence [3]. This is a vast field
with long tradition and includes, for example, various expert systems that seek answers from
databases based on a set of questions or adaptive sequence of questions. In recent times, the bulk of
attention within AI has focused on ML and deep learning systems (see Figure 1).
Various algorithms have been developed to implement ML behaviour. Each of these algorithms
provides a wide area of potential applications and, for any given problem, several algorithms may
provide a viable solution. Depending on what and how the ML algorithms “learn” they are divided in
three broad categories: supervised learning, unsupervised learning and reinforcement learning.
In supervised learning, the ML algorithm is provided with a (large) set of known data that corresponds
to the input in the problem domain and the desired output. This is akin to using open samples to train
humans. The data is called labelled because it contains the desired output (label) for included input
(inspection signal). The learning then proceeds to optimize the network to produce the desired output
when given any relevant input. The key benefit is that the desired outcome is clearly and explicitly
defined. A typical problem is that it is necessary to have a large set of labelled data, which is often
costly to produce or unavailable. For NDT the labelled data takes the form of NDT data from open
flawed and unflawed samples including e.g. position, type and orientation of each flaw.

ENIQ – European Network for Inspection & Qualification 1

 ENIQ RP13 – Qualification of NDT Systems that Make Use of ML

In unsupervised learning, the ML algorithm tries to optimize and create internal representation of the
data that follows the algorithm design. Such models can be used to find, e.g. anomalous input or items
closely matching a given input. The key benefit is that pre-labelled input data is not needed. Such
models could be taught to flag suspiciously “anomalous” signals even when known data from open
samples are not available.

Figure 1: A schematic overview of AI sub items [4]

In reinforcement learning, the ML algorithm takes active action and learns to optimize its actions to
maximize some “reward”, i.e. desired outcome, in a continuous development of itself. Such algorithms
can be used, e.g. for learning to play an interactive game, where the problem domain needs to be
“explored” to create the learning dataset.
While all of these ML models offer potential benefits in the field of NDT, supervised learning models
can be most easily integrated within the current ENIQ framework. The applicable models can be
trained with a controlled, verifiable training dataset and versioned (i.e. the learning frozen) so that an
immutable trained model can be qualified with predictable results. Chapter 4 focuses on such
supervised learning models while the Appendix provides some high-level descriptions of shallow and
deep architectures [5], and the associated methods that are widely deployed by the scientific
community.
Various schemes of continuous and incremental learning modes can also be envisioned. E.g. if it is
planned that the ML software uses reinforced learning after the qualification, this could potentially be
justified by freezing the qualified ML system (M1) and in parallel let a second ML system (M2) evolve
using new data collected during inspections. M2 evaluation results cannot be used during the
inspections but can be utilized in the next qualified revision of the software. Such continuous learning
schemes are excluded from the scope of this document.

ENIQ – European Network for Inspection & Qualification 2

 ENIQ RP13 – Qualification of NDT Systems that Make Use of ML

Analysis of NDT data can be divided in three tasks that need to be solved:
1. Detection needs to be assured by proper criteria for the identification, i.e. false calls excluded.
2. Characterization reveals the nature of the indication (e.g. orientation, surface breaking /
embedded, volumetric / planar) within the limitations of the inspection technique.
3. Sizing of indications provides the possibility of further assessment, e.g. fracture mechanics.

3.2. Cases of Use and Extent of Qualification

Data analysis models developed by ML algorithms can be applied with different approaches. The
qualification needs to be seen in conjunction with the responsibility and role of the human analyst and
for which task the model is supposed to be used. The following cases of use are discussed in terms of
the expected extent of qualification:
(1) Primary and classical analysis by humans is independently screened by a ML model.
Obviously, the most conservative approach is using a ML model as a secondary analysis
instance. Detection, positioning, characterization and/or sizing could be executed by a ML
model. This might reveal false calls of flaws or confirm the classical analysis result. It could
be concluded that qualification of the ML model is not required since classical analysis is
still performed. On the other hand, it should be considered that humans could be
influenced and biased by the ML model results, although the primary and secondary
analysis is meant to be independent. This would blur the boundaries to the following case
of use and would then require qualification.
(2) Classical analysis by humans is supported by a ML model.
The human analyses the complete dataset, but is alerted in areas where flaw-like patterns
are identified by the ML model. The responsibility can be ascribed to the human analyst.
Nevertheless, qualification of the ML model is necessary since the human analyst is
influenced by the ML model result.
(3) ML model substitutes part of the analysis by humans.
The human analysis is reduced to areas identified by the ML model as suspicious. Human
analysis of the entire data would be possible but not necessary and therefore the whole
process is more time efficient. The qualification of the ML model needs to focus on the
detection and positioning of flaws, since the responsibility for areas without identified
indications is taken over by the ML model. If the ML model performs characterisation
and/or sizing then qualification of these is necessary since the human analyst may be
influenced.
Another example for a substitute is a partly autonomous analysis by the ML model
without human interaction or decision in cases where humans’ logical reasoning or
experience cannot accomplish the task. This might be e.g. a data processing prior to
further human analysis. The impact on detection, positioning, characterization and sizing
is supposed to be beneficial for the overall performance. This requires qualification of all
aspects that are influenced by the task accomplished by the ML model.
(4) Completely autonomous analysis by the ML model.
This requires qualification of all aspects of data analysis: detection, positioning,
characterization and sizing.
The above considerations lead to the conclusion that qualification is necessary no matter how a ML
model is used.

ENIQ – European Network for Inspection & Qualification 3

 ENIQ RP13 – Qualification of NDT Systems that Make Use of ML

3.3. Challenge and Opportunity of Machine Learning

For a traditional (non-ML) data analysis the procedure describes analysis criteria, which were
determined during the development and demonstration process of the NDT system and is justified
using physical reasoning. Additionally, the experience of inspection personnel provides useful and
meaningful criteria for detection, positioning, characterization and sizing. Not all of them can be
entirely described in a procedure document. Beside the procedure, the data analyst still needs
experience for a successful qualification and examination.
The relevant parameters of flaws can be directly observed and suitable test flaws can be justified. In a
traditional qualification, it is often possible to cover the scope of a procedure with a limited number
of flaws concentrated around worst-case defects, together with the technical justification (TJ)
justifying the capability over the complete defect range via physical reasoning and modelling.
Experience and open literature with ML in the NDT community is so far limited. Training of the ML
system that provides the analysis criteria may therefore be more obscure and challenging to accept
than the experience of the human inspectors we are used to. Thus, more evidence may be required
for the ML system in the TJ, as detailed in Section 4.2. In addition, the worst-case defects used in
traditional qualifications may need to be adjusted for the ML system.
At the same time, ML systems provide considerable opportunity for improving the performance and
reliability of ISI. It is well known that even the highly experienced and qualified inspectors exhibit
variations in performance due to “human factors” [6] [7] [8]. Although much care is taken to make the
qualification trials as close to the real inspections as possible, there are necessarily differences that
may affect the apparent inspector performance [9].
In contrast, ML systems are expected to provide high repeatability and consistency in their application.
Their performance is directly observable in the qualification and expected to carry on unaltered in the
actual inspection setting.

ENIQ – European Network for Inspection & Qualification 4

 ENIQ RP13 – Qualification of NDT Systems that Make Use of ML

4. Qualifying Machine Learning NDT Systems

4.1. Example of ML Qualification Flow Chart
The flow chart in Figure 2 highlights the areas in the qualification process where ML has an impact.
Design of Inspection
Qualification Development of Design of open and
Defect description solution and
need qualification blind test blocks and
and detection target qualification
identified strategy ML data sets
concept

Development of Development of
Development of Creation of Selection and
Technical inspection
inspection and Qualification training of Machine
Justification for the technique for
analysis procedure dataset for ML Learning Software
Inspection system detection and sizing

Procedure &
Procedure Blind qualification Qualification
ML software
validation for human analysts approved
qualification

Normal Qualification
New activity due to
Qualification process affected by
ML
process ML

Figure 2: Flow chart highlighting areas in the qualification process where ML has an impact

4.2. Demonstrating the Performance of a ML System

The capability of a ML system is demonstrated in the same manner as for a conventional NDT system
using the ENIQ methodology, by a TJ and practical qualification trial. The justification of the capability
of the ML elements of the NDT system should be incorporated into the main TJ, although a separate
TJ could be produced specifically for the ML elements if required.
It is expected that ML is used in the context of data analysis and the guidelines below are written
accordingly.

4.2.1. Specific Aspects for the Technical Justification

The TJ for the NDT system should follow the structure and guidelines of ENIQ RP2 [10] and include
justification of the acquisition, human data analysis and ML sub-systems. ML should be considered in
the following sections of the TJ:
Section 3: Overview of the NDT System
The NDT system should be presented, detailing the role of the ML system. A summary of the ML system
should be given including the reasoning for its use, chosen algorithms and expected results.
Section 4: Analysis of the Influential Parameters
Due to the complexity of any ML system, it is not expected that every parameter is defined and justified.
Once the ML system has been fully commissioned and frozen, it can be treated as any other software
package. Key parameters along with the software version number and management of revision should
be included.

ENIQ – European Network for Inspection & Qualification 5

 ENIQ RP13 – Qualification of NDT Systems that Make Use of ML

It is anticipated that the justification for a ML system will rely more heavily on experimental evidence
than physical reasoning. Thus, the evidence to justify the influential parameters should be sufficient in
scope and depth to understand how it impacts the ML decisions concerning the inspection outcome.
ML related parameters to include, but not limited to, are:
• Choice of algorithm
While many different algorithms may provide a solution for a given inspection case, they may
differ in their potential failure modes and the needed justification. The chosen algorithm
should be introduced in sufficient detail to allow assessing the potential failure modes and
needed justification.
• Dataset
The quality and size of the dataset used in training and validating the ML model is crucial for
overall performance. Thus, the used dataset should be described in sufficient detail to allow
assessment of its sufficiency. In particular, the dataset should
• Cover the inspection target laid out in the input information,
• Contain representative flaws and representative non-flaws to sufficient extent and
• Not include any bias that could adversely affect the inspection results.
It is expected that data scarcity will be one of the key issues in ML development and various
data-augmentation schemes and simulated data may be used to expand the training data. Any
use of such schemes should be presented.
During the development of a ML system, the data is commonly divided into:
1. Training datasets used directly for training the ML model,
2. Test datasets used to monitor and evaluate the training, and
3. Qualification datasets used to evaluate the final performance of the model.
The selection and independence of these datasets have significant effect on the ML
performance and should be considered an influential parameter.
Section 5: Physical Reasoning
It is expected that most if not all the essential parameters detailed in Section 4 can be justified by
experimental evidence (Section 7). Physical reasoning may be used to justify the use of simulated
signals or other data-augmentation schemes.
It is expected that the ML system as a whole can be justified later in Section 9 as the ML element of
the NDT system can be treated as inspection software.
Section 7: Experimental Evidence
It is expected that the justification of performance will require use of suitable statistical performance
metrics, but this will depend on the system and the way ML is applied. The performance metrics should
be selected and justified to suit the particular system to be qualified. The metrics may include hit/miss
probability of detection (POD) [11], receiver operating characteristics (ROC) [12] or some other
suitable metric.
The confidence of the performance estimates should be assessed. This can be included in the standard
methodology applied (e.g. POD) or may be specific to the justification. The performance of a ML model
can be assessed by considering the uncertainties linked to the model predictions (model variability).
Such uncertainties (also known as epistemic uncertainty) are due to the model variability fit on the
provided dataset. Changing the training dataset may impact the ML model online performance. The

ENIQ – European Network for Inspection & Qualification 6

 ENIQ RP13 – Qualification of NDT Systems that Make Use of ML

estimation of this variability can be provided either by using probabilistic ML models or by evaluating
different ML models for a given training dataset (e.g., by cross-validation procedure).
Qualification criteria are normally focused on correct detection or characterization/sizing. False call
performance often receives secondary attention. In the case of a ML system, the false call rate may
serve as a leading indicator of potential performance problems. If an inspection involving ML is applied
to data that was not sufficiently represented in the training data, the performance is expected to
deteriorate. Since, in the case of NDT, the opportunities of making false calls are much more prevalent
than opportunities for making missed calls, it can be expected that deteriorating performance is first
observed as an increase in the false call rate. Thus, the false call rate can be used as an early warning
signal of a ML system and it may be advisable to screen the ML false call rate closely.
A ML system may be susceptible to a failure mode, where the model performs well on the training and
validation data used during development, but fails to generalize properly to unseen defects. This is
called overfitting and the absence of overfitting needs to be justified, typically by testing the ML model
on previously unseen data.
Section 8: Parametric Studies
Parametric studies may be used, for example, to give further justification for used data-augmentation
schemes and to demonstrate the representativeness of simulated data.
Section 9: Equipment, Data Analysis and Personnel Requirements
The ML system should be detailed in this section. Justification should be provided for the influential
and essential parameters defined in Section 4. Evidence should be provided on how the system has
been commissioned and version controlled.
Section 10: Review of Evidence Presented
A summary should be provided detailing the capability of the ML system demonstrated in the previous
sections.
Section 11: Conclusions and Recommendations
The capability of the ML system should be stated with any recommendations for improvement if the
full specification could not be achieved.

4.2.2. Qualification Practical Trials

Traditionally, practical trials are conventionally divided into open trials to validate the procedure
operations and blind trials to confirm operator performance. The open trial samples may contain flaws
smaller than the detection target to gain further confidence on the method performance.
In case of a ML system, the same basic structure is followed. However, the following changes and
specific practices are recommended.
For a ML system, the separation into “procedure performance” and “operator performance” is less
significant since the role of the operator is diminished. Consequently, it is recommended that the open
and blind practical trials are combined into a single blind trial which serves both purposes. Thus, the
practical procedure trial should be blind in the sense that the ML system has no prior information
about the datasets that will be used or the details of the performance of the system. This is important
in order to keep the blind trials truly independent of the model development, previous justification
and the TJ. The practical trials should nevertheless contain flaws designed to test the “procedures
capability”, similar to conventional open trials. The trial data may, e.g., contain flaws smaller than the
detection target to allow the qualification (QB) to assess “performance margin” with respect to the set
detection target, even if the detection of these flaws is not required.

ENIQ – European Network for Inspection & Qualification 7

 ENIQ RP13 – Qualification of NDT Systems that Make Use of ML

The criteria used by the QB should be set for the purpose of performance check, as in a non-ML
qualification, and may include, e.g., 100% detection result of flaws greater than or equal to the
detection target, allowed root-mean squared (RMS) error and allowed maximum under/oversizing
limits for sizing and positioning.
The practical trials typically utilize worst-case defects, meaning that some of the flaws in the practical
trials are selected to be especially challenging, to gain additional confidence on the method
performance. In practice, these worst-case defects are typically those that result in poor detectability
due to physical reasons of the NDT system, and thus are not specific to human inspector judgement.
Such use of worst-case defects is also valid for ML systems.
The ML system may exhibit an additional failure mode, where an indication relevant to a human
inspector is nevertheless missed or miss-characterized by the ML system. This failure mode would
indicate overfitting. To exclude overfitting, the model should be tested on unseen data to assess its
performance. The blind practical trials serve as an important final check against overfitting. The
definition of worst-case defects for the ML system may introduce considerations that were previously
not included for NDT systems with manual data analysis. E.g. flaw types that were under-represented
in the training data may be considered worst-case defects for the ML system.
In traditional qualification, it is sometimes possible to refer to previous evidence in the TJ and to
complete qualification without practical trials. This may be the case for a ML system as well, e.g. when
the ML system has already been used in similar settings and it can be demonstrated that the data the
ML system receives has not changed.

5. Summary
ML is well suited to be utilized in all inspection techniques where data can be digitalized. ML systems
provide considerable opportunity for improving the performance and reliability of ISI, as even the
highly experienced and qualified inspectors exhibit variation in performance. In contrast, ML systems
are expected to provide high repeatability and consistency in their application. The ML performance is
directly observable during the qualification and expected to carry on unaltered in the actual inspection
setting.
For the QBs the main change versus conventional qualifications is the design and maintenance of
relevant qualification test blocks and datasets. ML systems may require additional flaws. At the same
time, it is recommended that ML systems are qualified using blind trials (in contrast to open and blind
for the traditional), which may reduce mock-up requirements.
Qualifications must be done utilizing frozen software (i.e. the learning has stopped) to avoid changes
in performance during and after the qualification. The main challenge of inspection vendors is the
development of a solid justification presenting the functionality of the frozen ML software.
For the actual inspectors using the qualified ML software, whether as an assisted analysis, detection
tool or other, there is no change in the use of the conventional automated analysis as both are
versioned software in the eyes of the operator.
Further information is available in industry journal articles relating to the state-of-the-art of ML
development, targeting the nuclear inspection industry (see [13] as an example).

ENIQ – European Network for Inspection & Qualification 8

 ENIQ RP13 – Qualification of NDT Systems that Make Use of ML

References
[1] The European Methodology for Qualification of Non-Destructive Testing – Issue 4, ENIQ report
no. 61, The NUGENIA Association, 2019.
[2] ENIQ Position Paper: Qualification of an Artificial Intelligence / Machine Learning Non-
Destructive Testing System, ENIQ report no. 64, the NUGENIA Association, 2020.
[3] Russell, S. J. and Norvig, P., 2016. Artificial intelligence: A Modern Approach, Third edition,
Pearson Education Limited.
[4] Goodfellow, I., Bengio, Y. and Courville, A., 2016. Deep learning. MIT press.
[5] Bengio, Y. and LeCun, Y., 2007. Scaling learning algorithms towards AI. In L. Bottou, O. Chapelle,
D. DeCoste, & J. Weston (Eds.), Large-scale kernel machines, MIT Press.
[6] McGrath, B., 2008. Programme for the assessment of NDT in industry, PANI 3. Retrieved from
www.hse.gov.uk/research/rrpdf/rr617.pdf on May 7, 2019.
[7] Bertovic, M., 2015. Human Factors in Non-Destructive Testing (NDT): Risks and Challenges of
Mechanised NDT. Technical University Berlin.
[8] Cumblidge, S., D’Agostino, A., Morrow, S., Franklin, C., & Hughes, N., 2017. Review of Human
Factors Research in Nondestructive Examination. Retrieved from
www.nrc.gov/docs/ML1705/ML17059D745.pdf.
[9] Virkkunen, I., Koskinen, T., & Jessen-Juhler, O. Virtual round robin – a new opportunity to study
NDT reliability. To be published.
[10] ENIQ Recommended Practice 2: Strategy and Recommended Contents for Technical Justifications
– Issue 3, ENIQ report no. 54, The NUGENIA Association, 2018.
[11] ASTM E2862-18, Standard Practice for Probability of Detection Analysis for Hit/Miss Data,
American Society for Testing of Materials (ASTM), 2018.
[12] Fawcett, T., 2005. An introduction to ROC analysis. Pattern Recognition Letters, 27, pp. 861-874.
[13] Miorelli, R., Skarlatos, A., and Reboud, C., 2019. A machine learning approach for classification
tasks of ECT signals in steam generator tubes nearby support plate. AIP Conference –
Proceedings, 2102(1) , 090004, https://doi.org/10.1063/1.5099822 .
[14] Bishop, C. M., 2006. Pattern recognition and machine learning. Springer.
[15] Shawe-Taylor, J. and Cristianini, N., 2004. Kernel methods for pattern analysis. Cambridge
University Press.
[16] Cristianini, N. and Shawe-Taylor, J., 2004. Support Vector Machines and other kernel-based
learning methods. Cambridge University Press.
[17] Scholkopf, B. and Smola, A. J., 2001. Learning with kernels: support vector machines,
regularization, optimization, and beyond. MIT press.
[18] Salucci, M., Anselmi, N., Oliveri, G., Calmon, P., Miorelli, R., Reboud, C., & Massa, A., 2016. Real-
time NDT-NDE through an innovative adaptive partial least squares SVR inversion approach. IEEE
Transactions on Geoscience and Remote Sensing, 54(11), pp. 6818-6832.
[19] Ahmed, S., Reboud, C., Lhuillier, P. E., Calmon, P., & Miorelli, R., 2019. An adaptive sampling
strategy for quasi real time crack characterization on eddy current testing signals. NDT & E
International, 103, pp. 154-165.
[20] Virkkunen, I., Koskinen, T., Jessen-Juhler, O., & Rinta-Aho, J., 2021. Augmented Ultrasonic Data
for Machine Learning. Journal of Nondestructive Evaluation, 40(1), pp. 1-11.

ENIQ – European Network for Inspection & Qualification 9

 ENIQ RP13 – Qualification of NDT Systems that Make Use of ML

[21] Meng, M., Chua, Y. J., Wouterson, E., & Ong, C. P. K., 2017. Ultrasonic signal classification and
imaging system for composite materials via deep convolutional neural networks.
Neurocomputing, 257, pp. 128-135.
[22] Zhu, P., Cheng, Y., Banerjee, P., Tamburrino, A., & Deng, Y., 2019. A novel machine learning
model for eddy current testing with uncertainty. NDT & E International, 101, pp. 104-112.
[23] Munir, N., Kim, H.-J., Song, S.-J., & Kang, S.-S., 2018. Investigation of deep neural network with
drop out for ultrasonic flaw classification in weldments. Journal of Mechanical Science and
Technology, 32(7), pp. 3073-3080.
[24] Tang, Y., 2013. Deep learning using linear support vector machines. arXiv preprint
arXiv:1306.0239.
[25] Theodoridis, S., 2015. Machine learning: a Bayesian and optimization perspective. Academic
Press.
[26] ENIQ Glossary of Terms – Issue 3, ENIQ report no. 62, The NUGENIA Association, 2019.

ENIQ – European Network for Inspection & Qualification 10

 ENIQ RP13 – Qualification of NDT Systems that Make Use of ML

Appendix: Machine Learning Fundamentals

The Machine Learning Stages
Generally speaking, ML algorithms consist of the following stages:
• Stage 1, Data gathering: A set of data (labelled or not) is collected. The source of the data can
be of various origins (e.g., numerical calculations, experimental databases of NDT signals,
database of categories, etc.) depending on the task. The quality and the quantity of the training
data is crucial for the performance of the ML model.
• Stage 2, Data preparation: The data is formatted and cleaned to remove spurious content that
has meaningless information from a physics point of view. At this stage, data and signal
processing procedures can be applied as required. These may include e.g. image registration
or under sampling.
• Stage 3, Model choice or training phase: Once the data has been gathered and prepared, the
ML model can be fitted to the data. This phase is known as the training phase and is performed
until the model converges towards the stable solution. The training phase is based on a given
set of data (i.e., the training dataset) and is referred to as the off-line phase. During this stage,
a suitable set of model parameters is established based on the data provided to the model.
The obtained outcome at the end of the process is the trained ML model.
• Stage 4, Model evaluation or test phase: Once the model has been trained it can be tested on
a set of data that was not considered in the training phase. This stage is known as the test
phase. It does not require any learning and is therefore also known as the on-line phase. The
main purpose of this stage is to assess the performance of the ML model before deploying it.
The input data used for this stage should be collected and processed following the steps
defined in Stages 1 and 2.
• Stage 5, Prediction or inference: The final trained model is deployed. As for Stage 4, the input
data should be collected and prepared following the steps defined in Stages 1 and 2.

Shallow architectures
Shallow architecture refers to the set of ML methods that exploit the concept of kernel machines (KMs)
[14] [15] [16] [17]. In most cases, a supervised learning framework is used to perform classification
(e.g., flaw(s)/anomaly detection, defect(s) classification) or regression tasks (parametric-flaw
characterization). In order to deal with the cardinality of NDT raw signals, very often the KMs are coped
with a dimensionality reduction stage (part of data preparation phase) aiming at shrinking the
information content by reducing the redundancy on the learning signals and mitigate the so-called
curse-of-dimensionality issue [18]. This stage relies on well-established algorithms in many
communities (e.g., chemometric, signal processing, biomedical, etc.) from which one can establish
statistical and/or geometrical properties associated with the data reduction stage. Loosely speaking,
these methods can be grouped into two big families, the matrix decomposition algorithms such as
principal component analysis, independent component analysis etc., and the manifold learning
families of algorithms such as ISOMAP1, locally linear embedding, etc.
Different learning methods can be collected under the name of KMs. All KM methods rely on the use
of the so-called kernel trick [14] [16] [17] in order to perform classification and regression tasks. The
kernel trick enables linear interpolations of non-linear data by fitting the model directly in the kernel

1
ISOMAP is a nonlinear dimensionality reduction method.

ENIQ – European Network for Inspection & Qualification 11

 ENIQ RP13 – Qualification of NDT Systems that Make Use of ML

space [15]. For the sake of brevity, we mention just hereafter the most known and studied methods
by providing a brief and concise applicative background associated to each of them.
It is worth mentioning that the most widely deployed KMs in the literature repose on the vectorization
of data (i.e., the matrices or tensors are reshaped into vectors). Therefore, any kind of spatial and/or
temporal coherence within probed data is not preserved in the training and testing dataset.
The kernel ridge regression (KRR) is the kernel version of the well-known ridge regression [17]. KRR is
obtained by formulating the ridge regression exploiting the kernel trick. KRR enables the control on
the regression performance through a regularization coefficient that can be tuned in order to maximize
the trade-off between variance and bias. Due to its statistical meaning, this penalization coefficient
can be a valuable degree of freedom to enhance the model performance in case of noisy
measurements. This hyper-parameter, plus the ones associated to the chosen kernel (often just one)
are the only parameters to be estimated in order to obtain a classification or regression (e.g., defect
localization/characterization) model. The tuning of KRR hyper-parameters is often obtained via cross
validation.
The Gaussian Process (GP) for classification and regression, also known under the name of Kriging in
the geoscience community, is a statistical model that exploits the Bayesian framework in order to
perform classification and regression tasks [14]. Even though the GP formulation shares many common
points with KRR, its statistics enables the access to the mean and the variance of the prediction,
providing de facto a measure of the model or epistemic uncertainties associated to the predictions
(i.e., the classification or regression results). GP as KRR requires the tuning of hyper-parameters
associated to both the deployed kernel (i.e., the covariance function) and the regularization coefficient.
This stage is often performed via minimization of a suitable likelihood function.
Support Vector Machines (SVMs) for regression and classification tasks are widely and successfully
deployed in many different fields from late nineties. SVMs are based on a mathematical background
rooted in the statistical learning theory and structural risk minimization. It allows for theoretical limits
to be applied to the SVM model [16][17]. As all the other kernel methods, a SVM model requires tuning
of the kernel hyper-parameters along with two parameters associated to the SVM algorithms. The
physical meaning of SVM parameters and thus their choice, can be provided by the theory upon which
the SVM is developed. The tuning of SVM hyper-parameters is often obtained via cross validation.
Compared to KRR and GP, SVM model enables sparse predictions that turns into a computationally
efficient model compare to the above-mentioned methods [19].
It is worth mentioning that other shallow architectures have been developed in the literature. All these
models can be considered as improvements, modification or hybridization of the aforementioned ones.
The most known ones are co-kriging, universal-kriging and relevance vector machine, etc.

Deep Learning Methods

Deep learning (DL) methods have shown their potential in the first decade of 21 st century when they
appeared to be able to provide the same or better performance than shallow architectures applied to
supervised learning tasks in image classification problems. DL methods rely on the use of specific
neural network architectures like multilayer perceptron, convolutional neural network, etc. More
recently, the use of DL has been boosted by the increasing performances of Graphical Process Units
(GPUs) enabling more efficient model training. Research in DL have been motivated by the fact that DL
methods aim to avoid feature engineering and kernel engineering stages that are often necessary
before training a kernel machine model. This makes DL an attractive tool to solve regression and
classification problems for end-users that are non-experts in ML [4].
In neural networks, the learned function is formed by linear combination of a set of simple non-linear
activation functions. The model is typically formed in layers, where the layers are connected through
activation functions and results within a layer are combined linearly with learned weights.

ENIQ – European Network for Inspection & Qualification 12

 ENIQ RP13 – Qualification of NDT Systems that Make Use of ML

During learning, the input data is propagated through each layer to form the model result. This is then
compared to the given label value and an error value computed using a specified function (the “cost
function”). This error is then propagated backwards (back propagated) through the model and at each
layer the weights are updated to improve the next prediction. With each iteration, the model weights
of the whole network are updated to give, presumably, a better prediction.
For many problems, such as image classification, the location of the features sought are
inconsequential. This location-invariance can be introduced to the model by a convolution layer, where
a small kernel is shifted through the data and a value is computed at each location. The layers form
“feature maps” that encode location-invariant information about the presence and relationships of
specified features. This convolution effectively creates weight sharing that significantly reduces the
number of learned parameters. The number of learned parameters can be further reduced by pooling
layers, which combine activations for adjacent locations.
Deep convolutional neural networks (DCNNs) make use of convolutional and pooling layers, to encode
source data (often an image) to increasingly abstract representations while reducing the
dimensionality of the data with each subsequent layer. Such very deep models have proven very
successful in many image classification tasks.
Virkkunen et al. [20] used DCNNs to successfully detect cracks in phased array ultrasonic data. Recently
Meng et al. [21], Zhu et al. [22] and Munir et al. [23] used DCNNs for defect classification in ultrasonic
and EC-data, respectively. In general, the DCNNs are interesting for various flaw detection and
classification tasks and various NDT signal data.

Hybrid Learning by Coupling Shallow and Deep Architecture Methods

Coupling between shallow and deep architectures is a valuable solution in order to fully take advantage
of the solid mathematical background of KMs and the non-invasive (i.e., no-feature engineering stage
required) associated with DL approaches. This kind of hybrid learning approach is widely studied and
exploited by the ML community. It has been shown that hybrid learning approaches perform better
than KMs and DL methods for complex classification tasks [24].

Decision Trees
Decision trees (DTs) for classification and regression are non-parametric supervised methods that can
be applied to classification or regression problems [25]. The learning model is created by inferring a
decision rule based on data features. The decision chain is started from the root containing all features
and by successively splitting them while moving to the root children (called leaf node). The algorithm
stops when the whole tree depth is explored and convergence is achieved. Classification and
Regression Trees (CART) are the most widely used DT algorithms, which may also be used for sensitivity
analysis purposes.
The main strength of DTs consists in the fact that decisions are intuitive and easy to interpret (one can
visualize the decision graph). For this reason they are considered as white box models. Furthermore,
there is very little need to prepare data (scaling features) and the algorithm scales logarithmically with
respect to the training dataset size and can handle numerical and categorical data. The main
disadvantages of DTs are that they are prone to overfitting and that they are unstable with respect to
small variations in input data (a completely different tree can be obtained). These disadvantages can
be mitigated by deploying a ML method known as ensemble learning.

ENIQ – European Network for Inspection & Qualification 13

 ENIQ RP13 – Qualification of NDT Systems that Make Use of ML

Ensemble Learning
Ensemble learning (EL) is based on the concept that the aggregation of different ML methods
(classifiers or regressors) may give better prediction performance [25]. A group of predictions is
referred to as an ensemble, thus this technique is referred as ensemble learning or method.
Within EL different and possibly heterogeneous ML algorithms can be combined together. E.g. to
mitigate the disadvantages of DTs, one can combine different DTs together based on different input
subsets in order to obtain the prediction of all the DTs together. Such an algorithm is known as random
forest. In the recent past the ML scientific community has developed different EL algorithms, most
notably voting, bagging, boosting, stacking, etc. [25].

Design of Datasets
In the last decades ML research community have developed many ML models aiming to resolving
different kind of tasks ranging from image recognition, outlier detection, speech recognition, etc.
Generally speaking, among the best performer ML methods, one cannot a-priori infer that a ML model
with good performance in terms of accuracy and efficiency can be obtained for different problems.
This is also true for the NDT domain. Addressing ML problems, i.e., detection, classification or
regression, may require different ML models for each task. Furthermore, in order to obtain the desired
degree of performance for a given task, different ML models could be used for different inspection
methods and techniques. Evaluation of a batch of different candidate ML models is a robust and
feasible approach for a given inspection problem.
Apart from the choice of a suitable ML model, the availability of labelled data to train the ML model is
crucial. Indeed, the performance of a ML method depends significantly on the input data of the
algorithm. The input data of a supervised learning strategy is commonly referred to as training dataset.
The training dataset is composed of a set of input data (i.e., NDT probed signals) and output (or targets)
(e.g., flaw(s) categories, flaw detection, flaw(s) parameters estimation, etc.) data, which represent all
the information available to the ML method in order to establish a suitable model that links these two
datasets. In order to evaluate the ML model it must be tested for an unseen dataset referred to as a
test dataset. In order to evaluate the ML model performance, it is mandatory that the test dataset
must not contain any data points of the training dataset. In case of DL models, one should also account
for the creation of a validation dataset which is used to check the DL model accuracy in the training
phase.
In the context of NDT, ML methods can be trained based on both labelled experimental and simulated
data. The former is referred to as data-driven approach, the latter is referred to as physics-driven
approach, since the data is generated by a numerical model. A mix of mode- and data-driven
approaches can also be deployed. The generation of a training dataset based on a sub-set of simulated
signals can be seen as a viable way to add a-priori knowledge to the ML algorithm. In case of scarce
labelled experimental data, the synthetic data can improve the generalization capabilities of the ML
model and can give a model that is less prone to errors.

Training Dataset
One of the most important issues in developing a supervised ML system is the choice of the training
and test samples to properly fit and test the ML model. The training dataset should be representative
enough to contain a meaningful set of inspection configurations for a given inspection case (e.g.,
inspection of corrosion via eddy current testing, weld inspection via multi-element probe, etc.). The
suitable number of training samples depends on the deployed ML algorithm (shallow, deep
architectures) as well as on the ML model parameters. The correct number of training samples cannot
be fixed a-priori. Instead of fixing a given number of samples, one should demonstrate that a ML

ENIQ – European Network for Inspection & Qualification 14

 ENIQ RP13 – Qualification of NDT Systems that Make Use of ML

algorithm is able to increase its performance when more labelled input data is available. Doing so they
can infer on the convergence of the model with respect to the deployed training sample.
The choice of the training dataset has an impact on the choice of the ML model to be deployed.
Working with labelled data (i.e., supervised learning) enables access to classes of membership or to
continuous (physical) values. The ML model should be able to deal with different training datasets in
terms of input and target cardinality and volume, like e.g. prediction of discrete (i.e. classes) or
continuous, multi-variate values. The type of the targets impacts the performance of the ML model in
terms of accuracy training (and prediction) time. E.g. one can expect that working with A-scan, B-scan
or C-scan signals will not require the same computational effort due to the cardinality of the treated
problem. This implies that some ML models are more suitable than others.

Test Dataset
The test phase should contain the needed number of samples to check the robustness of the deployed
ML system. The test dataset should only be used to infer the potential of the ML model to
generalization capabilities (i.e., its robustness) on unseen test data. The experiments should be
representative for a variety of cases belonging to the same family of problems evaluated in the training
phase. To avoid bias in the obtained results, none of the test samples should be used in the training
phase. The number of test samples should be large enough to empirically provide statistical insight
(mean value and variance of predictions) of the obtained predictions.

Qualification Dataset
The qualification dataset has the same requirements as the test dataset, but with the caveat that this
dataset is designed and controlled by the qualification body (QB) and can only be used for qualification
purposes in order to avoid bias.

Guidelines on the Choice of ML Model and their Application in NDT

The performance of a ML model significantly depends upon the data to which the ML model is fitted.
For a given inspection method the generalization capability is linked to both the information content
of the training dataset and its cardinality. A given ML model may provide better performance than
another one depending on the data, i.e. the considered signals / measurements. In NDT, these signals
depend upon the inspection method and technique. In general we classify these signals as scalars, time
varying (A-scans, B-scans, and C-scans), 2D-images, 3D-images, etc. Such kind of signals may be found
in different inspection methods and techniques such as ultrasonic testing, eddy-current testing, etc.
Based on the application case the following recommendation can be given about ML-models to choose.
• Classification based on scalar signals (e.g., flaw(s) detection tasks): ML methods such as
shallow and deep architectures should provide thorough performance.
• Classification based on images-like signals (e.g., flaw(s) classification tasks): ML methods
such as shallow and deep architectures should provide thorough performance.
• Classification based on time-domain-like signals (e.g., flaw(s) classification tasks): Off-
the-shelf DL architectures may have an edge compared to traditional shallow architectures
for which careful data preparation, i.e., feature engineering, may be needed
• Regression based on images-like signals (e.g., flaw(s) sizing tasks): ML methods such as
shallow and deep architectures should provide thorough performance. For high-
dimensional regression problems (e.g., more than fifteen dimensions) DL architectures
may provide an edge in performance due to their learning procedure.

ENIQ – European Network for Inspection & Qualification 15

 ENIQ RP13 – Qualification of NDT Systems that Make Use of ML

• Regression based on time-domain-like signals (e.g., flaw(s) sizing tasks): Off-the-shelf DL

architectures may have an edge compared to traditional shallow architectures for which a
careful data preparation, i.e., feature engineering, may be needed.
According to the research in the field of ML it can be concluded that DL architectures often required a
high amount of training data in order to achieve good accuracy levels. Therefor the choice between
shallow or DL architectures also depends on the amount of available training data for a given task.

Model Evaluation
Assessing the performance of a ML system is crucial. To address this issue the ML research community
has provided a wide set of metrics that can be used to quantitatively compare different ML systems.
Two major families of metrics can be distinguished with respect to the problem that needs to be solved,
regression or classification.
In case of a regression problem, one can rely on different error metrics that are commonly used to
evaluate the performance of a ML model. The most common metrics are provided in Error! Reference
source not found.. In case of classification along with an accuracy estimation, one would like to carry
out a study of the ML model performance itself. This concept is much related to the evaluation of the
performance of an inspection procedure via POD studies. For ML systems one rarely relies on scalar
inputs as for POD studies. Thus, the concept of detection threshold cannot be applied for flaw
detection in a straightforward manner. To address this issue tools like ROC curve, Precision/Recall
curves and other quantities that can be derivate from Fawcett [12] are used instead.
ROC curves have been developed to evaluate the detection performance in early radar systems in noisy
environments and have been applied for a long time in biomedical and ML domains to evaluate the
presence of possible diseases or pathologies. ROC curves are in particular useful in case of with
skewed/ unbalanced class distribution and to evaluate cost-sensitive learning. Such kind of problems
is quite common in real inspection for which very few anomalies are detected among many other signal
indications.

ENIQ – European Network for Inspection & Qualification 16

 ENIQ RP13 – Qualification of NDT Systems that Make Use of ML

Table 1: Common error metrics to assess the performance of a ML model for regression purposes

Error metric
𝑁
1
Mean-squared error MSE = ∑(𝑝𝑖 − 𝑎𝑖 )2
𝑁
𝑖=0

𝑁
1
Root mean-squared error RMSE = √ ∑(𝑝𝑖 − 𝑎𝑖 )2
𝑁
𝑖=0

𝑁
1
Mean-absolute error MAE = ∑|𝑝𝑖 − 𝑎𝑖 |
𝑁
𝑖=0
2
∑𝑁
𝑖=0(𝑝𝑖 − 𝑎𝑖 )
Relative-squared error* RSE = 2
∑𝑁
𝑖=0(𝑎𝑖 − 𝑎
̅)
2
∑𝑁 (𝑝𝑖 − 𝑎𝑖 )
Root relative-squared error* RRSE = √ 𝑖=0
∑𝑁 ̅)2
𝑖=0(𝑎𝑖 − 𝑎

∑𝑁
𝑖=0|𝑝𝑖 − 𝑎𝑖 |
Relative-absolute error* RAE = 𝑁
∑𝑖=0|𝑎𝑖 − 𝑎̅|
𝑁
∑𝑖=0(𝑎𝑖 − 𝑝𝑖 )2
Coefficient of determination** 2
R =1−
∑𝑁 ̅)2
𝑖=0(𝑎𝑖 − 𝑎
Where in (*) 𝑎̅ is the mean value over the training data and in (**) 𝑎̅ is over the test data. 𝑝𝑖 stands
for the i-th predicted instance and 𝑎𝑖 is its actual counterpart.

ENIQ – European Network for Inspection & Qualification 17

 ENIQ RP13 – Qualification of NDT Systems that Make Use of ML

Glossary
The general definitions in the ENIQ Glossary [26] are applicable to this RP. In addition, the following
definitions apply.
Artificial Intelligence Computer systems that mimic human intelligence or perform tasks
that have been thought to require human intelligence.
Machine learning (ML) method Algorithm to create and improve the performance of a model
through training data.
Machine learning (ML) model Programme for the prediction of a result based on input data. The
prediction can be related to regression or classification problems.
Training dataset A digital set of NDT data, representative of the inspection
configuration, used to develop and teach a ML system.
Test dataset A digital set of NDT data, representative of the inspection
configuration, kept independent to the training data and used to
verify functionality outside the training data.
Qualification dataset A digital set of NDT data, representative of the inspection
configuration, controlled and managed by the qualification body and
only used for the qualification of the ML system.

ENIQ – European Network for Inspection & Qualification 18

 ENIQ RP13 – Qualification of NDT Systems that Make Use of ML

Contributors to Drafting and Editing

Iikka Virkkunen Trueflaw Ltd. Finland

Martin Bolander Westinghouse Electric Sweden
Heikki Myöhänen Kiwa Inspecta Finland
Roberto Miorelli CEA France
Ola Johansson Swedish Qualification Centre (SQC) Sweden
Philip Kicherer Swiss Association for Technical Inspections (SVTI) Switzerland
Chris Curtis Jacobs / Inspection Validation Centre (IVC) Great Britain
Oliver Martin European Commission – Joint Research Centre European Commission

ENIQ – European Network for Inspection & Qualification 19

ABOUT ENIQ AND SNETP

The European Network for Inspection and Qualification (ENIQ) is a utility driven network working
mainly in the areas of qualification of non-destructive testing (NDT) systems and risk-informed in-
service inspection (RI-ISI) for nuclear power plants (NPPs). Since its establishment in 1992 ENIQ has
issued over 60 documents. Among them are the “European Methodology for the Qualification of Non-
Destructive Testing” and the “European Framework Document for Risk-Informed In-Service
Inspection”. ENIQ is recognised as one of the main contributors to today’s global qualification
guidelines for in-service inspection.
ENIQ is the technical area 8 of NUGENIA, one of the three pillars of the Sustainable Nuclear Energy
Technology Platform (SNETP) that was established in September 2007 as a R&D&I platform to support
technological development for enhancing safe and competitive nuclear fission in a climate-neutral
and sustainable energy mix. Since May 2019, SNETP has been operating as an international non-profit
association (INPA) under the Belgian law pursuing a networking and scientific goals. It is recognised as
a European Technology and Innovation Platform (ETIP) by the European Commission.
The international membership base of the platform includes industrial actors, research and
development organisations, academia, technical and safety organisations, SMEs as well as non-
governmental bodies.

secretariat@snetp.eu www.snetp.eu SNETP SNE_TP