J U L Y 2 0 2 0 • M A T E R I A L S E V A L U A T I O N 869 A B S T R A C T NDT 4.0 is a vision for the next generation of nondestructive inspection systems following the expected fourth industrial revolution based on connected cyber-physical systems. While an increasing use of automation and algorithms in nondestructive testing (NDT) is expected over time, NDT inspectors will still play a critical role in ensuring NDT 4.0 reliability. As a counterpoint to recent advances in artificial intelligence algo- rithms, intelligence augmentation (IA) refers to the effective use of information technology to enhance human intelligence. While attempting to replicate the human mind has encountered many obstacles, IA has a much longer history of practical success. This paper introduces a series of best practices for NDT IA to support NDT 4.0 initiatives. Algorithms clearly have a great potential to help alleviate the burden of “big data” in NDT however, it is important that inspectors are involved in necessary secondary indication review and the detection of rare event indications not addressed well by typical algorithms. Examples of transitioning algo- rithms for NDT applications will be presented, emphasizing the successful interfacing of inspector and software for optimal data review and decision making. KEYWORDS: Industry 4.0, artificial intelligence, intelligence augmentation, human-machine interface, reliability Introduction Industry 4.0 is a term developed by German industry leaders and researchers to describe how the Internet of Things (IoT), an emerging network of linked cyber-physical devices, will improve engineering, manufacturing, logistics, and life-cycle management processes (Jahanzaib and Jasperneite 2013). The number 4.0 refers to a fourth industrial revolution. Begin- ning in the 1700s, three major waves of technological changes transformed the industrial landscape and increased produc- tivity: (1) mechanization and water/steam power (2) mass production (for example, assembly lines) and electricity and (3) computers and automation. The fourth industrial revolu- tion is expected to be based on connected cyber-physical systems. There is a parallel vision for the next generation of NDT capability referred to as NDT 4.0 (Meyendorf et al. 2017a, 2017b Link and Riess 2018 Vrana et al. 2018 Singh 2019). A key aspect of NDT 4.0 is leveraging automation in the evaluation of the workpiece and providing characteriza- tion of the state of the part for improved life-cycle manage- ment (Lindgren 2017 Forsyth et al. 2018). A diagram of an integrated vision for NDT 4.0 is presented in Figure 1. One key innovation of NDT 4.0 is the integration of advanced control systems and NDT algorithms to support complex inspections, NDT sensor data acquisition, and data analysis tasks. In recent years, major advances have been made in the field of machine learning and artificial intel- ligence (AI) to perform complex data classification tasks, leveraging training on “big data” sets (LeCun et al. 2015). While this technology is promising, challenges do exist with transitioning machine learning/AI algorithms to NDT appli- cations. Training AI requires very large, well-understood data sets, frequently not available in NDT, and there are major concerns about the reliability and adaptability of such algo- rithms to completely perform complex NDT data review tasks. One of the primary objectives of this paper is to survey the potential benefits and challenges of emerging algorithms for NDT 4.0 systems. Experience and perspective on the tran- sition of algorithms for NDT applications will also be discussed. The Inspector and NDT 4.0 A critical component of any NDT tool is the interface with the human that uses it. Figure 1 shows the human-machine interface as a critical link between the NDT inspector/engineer and Intelligence Augmentation and Human-Machine Interface Best Practices for NDT 4.0 Reliability by John C. Aldrin* * Computational Tools, Gurnee, Illinois 60031, USA aldrin@computationaltools.com Materials Evaluation 78 (7): 869–879 https://doi.org/10.32548/2020.me-04133 ©2020 American Society for Nondestructive Testing

870 M A T E R I A L S E V A L U A T I O N • J U L Y 2 0 2 0 NDT 4.0 software and hardware. Care must be taken with the implementation of automation to ensure that operators have the necessary awareness and control as needed. In addition, as a counterpoint to recent advances in AI algorithms, intelli- gence augmentation (IA) is introduced as the effective use of information technology to enhance human intelligence. From this perspective, the inspector is an integrated part of NDT 4.0 systems and performs necessary tasks in collaboration with automated NDT systems and data analysis algorithms. This paper will present a series of best practices for the interface between NDT hardware, software, and algorithms and human inspectors and engineers to ensure NDT 4.0 reliability. Algorithms and AI in NDT NDT algorithms that perform indication detection and char- acterization can be organized into three classes: (1) algorithms based on NDT expert knowledge and procedures (heuristic algorithms) (2) model-based inversion and (3) algorithms incorporating statistical classifiers and/or machine learning. The most basic algorithm is one based on human experience. The term heuristic algorithm is useful to describe a class of algorithms based on learning through discovery and incorpo- rating rules of thumb, common sense, and practical knowl- edge. This first class of algorithms essentially encodes all key evaluation steps and criteria used by operators as part of a procedure into the algorithm. The second class of algorithms is a model-based inversion that uses a “first principles” physics-based model with an iterative scheme to solve char- acterization problems. This approach requires accurate forward models and iteratively compares the simulated and measurement data, adjusting the model parameters until agreement is reached. The third class of algorithms covers statistical classifiers and machine learning, which are built through the fitting of a model function using measurement “training” data with known states. Statistical representation of data classes can be accomplished using either frequentist procedures or Bayesian classification. Machine learning and AI are general terms for the process by which computer programs can learn. Early work on machine learning built upon emulating neurons through functions as artificial neural networks using layered algorithms and a training process that mimics a network of neurons (Fukushima and Miyake 1982). In recent years, impressive advances have been made in the field of machine learning, primarily through significant developments in deep learning neural network (DLNN) algorithms (Hinton et al. 2006 LeCun et al. 2015 Lewis-Kraus 2016). Large sets of high-quality, well-characterized data have been critical for the successful training of DLNNs. As well, software tools have been devel- oped for training neural networks that better leverage advances in high-performance computing. A recent ME TECHNICAL PAPER w ia and human-machine interfaces NDT inspector/ engineer Test article NDT algorithms and models (artificial intelligence, digital twin) Human-machine interface (HMI) NDT automation control (scan plans) NDT automation hardware (scanning, cobots, drones) NDT sensors and data acquisition (SHM/IoT, databases) NDT 4.0 intelligence augmentation (IA) NDT 4.0 software NDT 4.0 hardware Figure 1. Vision for NDT 4.0: intelligence augmentation (IA) for NDT inspectors and engineers is achieved through a human-machine interface to NDT automation hardware, sensors, and data acquisition algorithms and models.