Domain Adapted Deep-Learning for Improved Ultrasonic Crack Characterization Using Limited Experimental Data.

Deep learning is an effective method for ultrasonic crack characterization due to its high level of automation and accuracy. Simulating the training set has been shown to be an effective method of circumventing the lack of experimental data common to nondestructive evaluation (NDE) applications. However, a simulation can neither be completely accurate nor capture all variability present in the real inspection. This means that the experimental and simulated data will be from different (but related) distributions, leading to inaccuracy when a deep learning algorithm trained on simulated data is applied to experimental measurements. This article aims to tackle this problem through the use of domain adaptation (DA). A convolutional neural network (CNN) is used to predict the depth of surface-breaking defects, with in-line pipe inspection as the targeted application. Three DA methods across varying sizes of experimental training data are compared to two non-DA methods as a baseline. The performance of the methods tested is evaluated by sizing 15 experimental notches of length (1-5 mm) and inclined at angles of up to 20° from the vertical. Experimental training sets are formed with between 1 and 15 notches. Of the DA methods investigated, an adversarial approach is found to be the most effective way to use the limited experimental training data. With this method, and only three notches, the resulting network gives a root-mean-square error (RMSE) in sizing of 0.5 ± 0.037 mm, whereas with only experimental data the RMSE is 1.5 ± 0.13 mm and with only simulated data it is 0.64 ± 0.044 mm.

Deep Learning for Ultrasonic Crack Characterization in NDE.

Machine learning for nondestructive evaluation (NDE) has the potential to bring significant improvements in defect characterization accuracy due to its effectiveness in pattern recognition problems. However, the application of modern machine learning methods to NDE has been obstructed by the scarcity of real defect data to train on. This article demonstrates how an efficient, hybrid finite element (FE) and ray-based simulation can be used to train a convolutional neural network (CNN) to characterize real defects. To demonstrate this methodology, an inline pipe inspection application is considered. This uses four plane wave images from two arrays and is applied to the characterization of cracks of length 1-5 mm and inclined at angles of up to 20° from the vertical. A standard image-based sizing technique, the 6-dB drop method, is used as a comparison point. For the 6-dB drop method, the average absolute error in length and angle prediction is ±1.1 mm and ±8.6°, respectively, while the CNN is almost four times more accurate at ±0.29 mm and ±2.9°. To demonstrate the adaptability of the deep learning approach, an error in sound speed estimation is included in the training and test set. With a maximum error of 10% in shear and longitudinal sound speed, the 6-dB drop method has an average error of ±1.5 mmm and ±12°, while the CNN has ±0.45 mm and ±3.0°. This demonstrates far superior crack characterization accuracy by using deep learning rather than traditional image-based sizing.

Fusion of multi-view ultrasonic data for increased detection performance in non-destructive evaluation.

State-of-the-art ultrasonic non-destructive evaluation (NDE) uses an array to rapidly generate multiple, information-rich views at each test position on a safety-critical component. However, the information for detecting potential defects is dispersed across views, and a typical inspection may involve thousands of test positions. Interpretation requires painstaking analysis by a skilled operator. In this paper, various methods for fusing multi-view data are developed. Compared with any one single view, all methods are shown to yield significant performance gains, which may be related to the general and edge cases for NDE. In the general case, a defect is clearly detectable in at least one individual view, but the view(s) depends on the defect location and orientation. Here, the performance gain from data fusion is mainly the result of the selective use of information from the most appropriate view(s) and fusion provides a means to substantially reduce operator burden. The edge cases are defects that cannot be reliably detected in any one individual view without false alarms. Here, certain fusion methods are shown to enable detection with reduced false alarms. In this context, fusion allows NDE capability to be extended with potential implications for the design and operation of engineering assets.

Data Fusion of Multiview Ultrasonic Imaging for Characterization of Large Defects.

The multiview total focusing method (TFM) enables a region of interest within a specimen to be imaged using different ray paths and wave-mode combinations. For defects larger than the ultrasonic wavelength, different portions of the same defect may manifest in a number of views. For a crack, the tip diffraction response may be evident in certain views and the specular reflection in others. Accurate characterization of large defects requires the information in multiple views to be combined. In this work, three data fusion methodologies are presented: a simple sum over all views, a sum weighted according to the inverse of the noise in each view, and a matched filter approach. Four large defects are examined; one stress corrosion crack (SCC), two weld cracks, and a pair of slagline defects in a weld. The matched filter (matched to a small circular void) provided significant improvement over the best individual view. The data fusion process incorporates artifact removal, where nondefect artifact signals within each image view are identified and masked, using a single defect-free data set for training. The matched filter was able to accurately visualize the full 3-D extent of the four defects, allowing characterization via the decibel drop method. When compared to X-ray computed tomography and micrograph data in the case of the SCC, the matched filter fusion provided excellent agreement. Its performance was also superior to any individual view while providing a single fused image that is easier for an operator to interpret than a set of multiview images.

A model for multi-view ultrasonic array inspection of small two-dimensional defects.

The multi-view total focusing method (TFM) is an imaging algorithm for ultrasonic full matrix array data that exploits internal reflections and mode conversions in the inspected object to create multiple images, the views. Modelling the defect response in multi-view TFM is an essential first step in developing new detection and characterisation methods which exploit the information present in these views. This paper describes a ray-based forward model for small two-dimensional defects and compares its results against finite-element simulations and experimental data for the inspection of a side-drilled hole, a notch and a crack. A simpler version of this model, based on a single-frequency approximation, is derived and compared. A good agreement with the multi-frequency model and a speed-up of several orders of magnitude are achieved.

Experimental Quantification of Noise in Linear Ultrasonic Imaging.

An efficient procedure for experimental-based quantification of statistical distributions of both the random and microstructural speckle noise within an ultrasonic image is presented. This is of particular interest in the multiview total focusing method, which enables many images (views) of the same region to be obtained by utilizing alternative ray paths and mode conversions. For example, in an immersion configuration, 21 separate views of the same region of a sample can be formed by exploiting direct and skip paths. These views can be combined through some form of data fusion algorithm to improve defect detection and characterization performance. However, the noise level is different in different views and this should be accounted for in any data fusion algorithm. It is shown that by using only one set of experimental data from a single measurement location, rather than numerous independent locations, it is possible to obtain accurate noise parameters at an imaging level. This is achieved by accounting for the spatial variation in the noise parameters within the image, due to beam spread, directivity, and attenuation with a simple empirical correction. An important feature of the process is the suppression of image artifacts caused by signal responses from other ray paths with the use of image masking. This masking process incorporates knowledge of the expected autocorrelation length (ACL) of image speckle noise and high-amplitude cluster suppression. The expected ACL is determined via a simple ray-based forward model of a single point scatterer. Compared to the estimates obtained using multiple independent locations, the speckle noise parameters estimated from a single measurement location were within 0.4 dB.