Flaw detection in non-destructive testing, especially for complex signals like ultrasonic data, has thus far relied heavily on the expertise and judgement of trained human inspectors. While automated systems have been used for a long time, these have mostly been limited to using simple decision automation, such as signal amplitude threshold. The recent advances in various machine learning algorithms have solved many similarly difficult classification problems, that have previously been considered intractable. For non-destructive testing, encouraging results have already been reported in the open literature, but the use of machine learning is still very limited in NDT applications in the field. Key issue hindering their use, is the limited availability of representative flawed data-sets to be used for training. In the present paper, we develop modern, deep convolutional network to detect flaws from phased-array ultrasonic data. We make extensive use of data augmentation to enhance the initially limited raw data and to aid learning. The data augmentation utilizes virtual flaws—a technique, that has successfully been used in training human inspectors and is soon to be used in nuclear inspection qualification. The results from the machine learning classifier are compared to human performance. We show, that using sophisticated data augmentation, modern deep learning networks can be trained to achieve human-level performance.

Modern ultrasonic inspections utilize ever-richer data-sets made possible by phased array equipment. A typical inspection may include tens of channels with different refraction angle, that are acquired at high speed. These rich data sets allow highly reliable and efficient inspection in complex cases, such as dissimilar metal or austenitic stainless steel welds. The rich data sets allow human inspectors to detect cracks with low signal-to-noise ratio from the wider signal patterns. There’s a clear trend in the industry to even richer data sets with full matrix capture (FMC) and related techniques. Convolutional neural networks have recently shown capability to detect flaws with human level accuracy in ultrasonic signals at the B-scan level. To enable automated flaw detection at human-level accuracy for critical applications, these neural networks need be developed to take advantage of today’s rich phased array data-sets. In the present paper, we extend previous work and develop convolutional neural networks that perform highly reliable flaw detection on typical multi-channel phased array data on austenitic welds. The results show, that the modern neural networks can accommodate the rich ultrasonic data and display high flaw detection performance.

No abstract

Previous research (Li et al., Understanding the disharmony between dropout and batch normalization by variance shift. CoRR abs/1801.05134 (2018). http://arxiv.org/abs/1801.05134arXiv:1801.05134) has shown the plausibility of using a modern deep convolutional neural network to detect flaws from phased-array ultrasonic data. This brings the repeatability and effectiveness of automated systems to complex ultrasonic signal evaluation, previously done exclusively by human inspectors. The major breakthrough was to use virtual flaws to generate ample flaw data for the teaching of the algorithm. This enabled the use of raw ultrasonic scan data for detection and to leverage some of the approaches used in machine learning for image recognition. Unlike traditional image recognition, training data for ultrasonic inspection is scarce. While virtual flaws allow us to broaden the data considerably, original flaws with proper flaw-size distribution are still required. This is of course the same for training human inspectors. The training of human inspectors is usually done with easily manufacturable flaws such as side-drilled holes and EDM notches. While the difference between these easily manufactured artificial flaws and real flaws is obvious, human inspectors still manage to train with them and perform well in real inspection scenarios. In the present work, we use a modern, deep convolutional neural network to detect flaws from phased-array ultrasonic data and compare the results achieved from different training data obtained from various artificial flaws. The model demonstrated good generalization capability toward flaw sizes larger than the original training data, and the effect of the minimum flaw size in the data set affects the $$a_{90/95}$$ a 90 / 95 value. This work also demonstrates how different artificial flaws, solidification cracks, EDM notch and simple simulated flaws generalize differently.

Flaw detection in non-destructive testing, especially in complex signals like ultrasonic data, has thus far relied heavily on the expertise and judgement of trained human inspectors. While automated systems have been used for a long time, these have mostly been limited to using simple decision automation, such as signal amplitude threshold.The recent advances in various machine learning algorithms have solved many similarly difficult classification problems, that have previously been considered intractable. For non-destructive testing, encouraging results have already been reported in the open literature, but the use of machine learning is still very limited in NDT applications in the field. Key issue hindering their use, is the limited availability of representative flawed data-sets to be used for training.In the present paper, we develop modern, very deep convolutional network to detect flaws from phased-array ultrasonic data. We make extensive use of data augmentation to enhance the initially limited raw data and to aid learning. The data augmentation utilizes virtual flaws -a technique, that has successfully been used in training human inspectors and is soon to be used in nuclear inspection qualification. The results from the machine learning classifier are compared to human performance. We show, that using sophisticated data augmentation, modern deep learning networks can be trained to achieve superhuman performance by significant margin.

Ultrasonic testing is an important tool for assessing the structural integrity of pressurised components in nuclear power plants during in-service inspection. Consequently, the reliability of inspection is of high significance. In particular, it is important to determine the largest crack that could conceivably be missed during in-service inspection. This information is utilised in order to choose the most effective method for different situations. Probability of detection (POD) curves are used to quantify the effectiveness of the inspection. However, these POD curves require a lot of data points in order to provide reliable estimates of the lower limit performance. Traditionally, obtaining these curves has been relatively expensive and this is why different simulation tools have been used to reduce costs and the number of physical test-pieces needed. In the current study, a novel method based on limited test-pieces is explored to produce a POD curve from the measured data. The present data contains just three artificial cracks made with thermal fatigue. This is insufficient to produce a POD curve. However, the data was sufficient to evaluate the novel approach and to gain preliminary estimates of the expected POD and the most important influential parameters. In this study, the idea is to emulate the amplitude response from the measured crack in a way that represents an amplitude response from a certain crack size. This amplitude data is converted to a B-scan image from which the inspector will evaluate whether there is a crack or not. Then, a POD curve is generated from the achieved hit/miss data.

For estimating the probability of detection (POD) in non-destructive evaluation (NDE), there are two standard methods, the so-called â versus a approach and the hit/miss approach. The two approaches have different requirements for the quality and quantity of input data as well as for the underlying NDE method. There is considerable overlap between the methods, and they have different limitations, so it is of interest to study the differences arising from using each methodology. In particular, if the dataset is not ideal, the methodologies may exhibit different problems dealing with various limitations in the data. In this paper, a comparison between â versus a and hit/miss analysis was completed for two different data sets, a manual aerospace eddy-current inspection and a nuclear industry phased array ultrasonic weld inspection using a simplified online tool. It was found that the two standard methods (â vs. a and hit/miss) may give significantly different results, if the true hit/miss decision is based on inspector judgement and not automated signal threshold. The true inspector hit/miss performance shows significant variance that is not attributable to signal amplitude. Model-assisted POD was not able to model the inspector performance due to lack of representative amplitude threshold and difficulties in capturing true signal variance. The paper presents experience from practical cases and may be considered a European viewpoint.

No abstract