Abstract

Nondestructive testing has become an essential part of the maintenance of modern gas turbine blades and vanes since it provides an increase in both safety against critical failure and efficiency of operation. Targeted repairs of the blade's airfoil require localized wall thickness information. This information, however, is hard to obtain by nondestructive testing due to the complex shapes of surfaces, cavities, and material characteristics. To address this problem, we introduce an automated nondestructive testing system that scans the part using an immersed ultrasonic array probe guided by a robot arm. For imaging, we adopt a two-step, surface-adaptive Total Focusing Method (TFM) approach. For each test position, the TFM allows us to identify the outer surface, followed by calculating an adaptive image of the interior of the part, where the inner surface's position and shape are obtained. To handle the large volumes of data, the surface features are automatically extracted from the TFM images using specialized image processing algorithms. Subsequently, the collection of 2D extracted surface data is merged and smoothed in 3D space to form the outer and inner surfaces, facilitating wall thickness evaluation. With this approach, representative zones on two gas turbine vanes were tested, and the reconstructed wall thickness values were evaluated via comparison with reference data from an optical scan. For the test zones on two turbine vanes, average errors ranging from 0.05 mm to 0.1 mm were identified, with a standard deviation of 0.06–0.16 mm.

Graphical Abstract Figure
Graphical Abstract Figure
Close modal

1 Introduction

1.1 Motivation.

Blades and vanes of modern, efficient gas turbines are high-tech parts in a demanding environment [1,2]. To ensure safety and efficiency of operation, these parts regularly undergo a maintenance, repair, and overhaul cycle (MRO). During targeted repair, especially on the airfoil, it is important to know its local thickness since dropping below critical thickness values would lead to having to abandon the part. The thickness of turbine blades is usually evaluated on several specific points across the surface using manual ultrasonic testing (UT) with single-element probes in contact technique [3]. This testing setup is relatively inaccurate in the presence of two-dimensional curvature and regions of non-parallel back walls. It requires not only specially trained but also experienced nondestructive testing (NDT) personnel. To obtain more reliable and consistent results, the process shall be automated.

1.2 Robotic Ultrasonic Testing.

While progress has been made in developing mechanized NDT applications for the last decades [4], the requirements of a task like the present one cannot be met with conventional manipulators due to the complex geometry. For large-scale test objects like storage tanks, vessels, or pipelines, the term robotic NDT refers to robots that are crawling, climbing, or swimming inside or outside of these structures [5] while performing more conventional NDT tasks on the local surface. In the course of industry 4.0, increasingly complex NDT tasks on human-scale test objects are automated using robotic arms. In UT, this type of robotic automation has been trialed for some applications: Several publications cover the challenge of robot path planning for NDT [614]. The key aspect here is to create robot paths with adjustable distances between test points and scan lines while orienting the probe perpendicular to the local surface. Often, these requirements go beyond conventional robot path planning; therefore, individual solutions are necessary. The NDT robot paths can be achieved using the CAD model or even by acquiring models of the test objects using optical scanning systems attached to the robot. Special NDT applications use dual robotic arms, e.g., for in-process weld testing [15] or testing of thin composite structures [16].

1.3 Robot-Guided Wall Thickness Evaluation.

Recently, Zhou et al. [17] proposed a method for pointwise robot-guided UT wall thickness evaluation for turbine blades using a stationary single-element probe. The method shows high accuracy and stability but is limited to a few dozen predefined test positions. The single-element probe method is, however, less suitable for areas where the back wall is not approximately parallel to the outer surface. To acquire a comprehensive wall thickness mapping of the test object, it is necessary to go beyond single-point testing. The back wall's local position and geometry must be evaluated in an aperture. Given the back wall's characteristic features in turbine blades, e.g., turbulators and arc-like structures that divide cavities, this requires an imaging method.

1.4 Ultrasonic Imaging.

With ultrasonic array imaging, the spatial resolution can be highly increased compared to single-element UT. Meanwhile, the number of testing positions can be kept low. There are different techniques of UT imaging: In impulse-echo configuration, spatial scanning with focused beam single-element probes can be used to produce B- and C-scans. For non-focusing probes, data from the different test positions can be merged using the Synthetic Aperture Focusing Technique (SAFT) [18,19], which is a standard in UT and can increase the spatial resolution with respect to B- and C-scans. Further, having multiple independent ultrasonic elements in one transducer, i.e., an array probe, allows transmitting and receiving UT signals from multiple positions simultaneously. Exciting the elements with specifically calculated delay times makes it possible to form individual ultrasonic beams by linear superposition. This technique is called phased array [2023] and is also a standard in UT. The ultrasonic pulse can be steered in the desired direction or focused on individual points. One step further, it can be utilized that linear superposition can also be applied in post-processing if the linear contributions of the transmit-receive combinations of the elements are provided. This leads to the modern imaging technique named the Total Focusing Method (TFM) [24,25], where for each image pixel, a sum of amplitudes from the ultrasonic data is calculated. The algorithm takes the amplitudes from the A-scans at corresponding times such that a virtual focus is created at each pixel. TFM is usually applied to data from contact UT and requires a Full Matrix Capture (FMC) dataset, which consists of the A-scans of all possible combinations of sending and receiving elements. There are several extensions of FMC-TFM with a focus on reducing calculation time [2628] and raising detection quality [2931] but also widening the scope of applicability, e.g., for immersion testing [32] or anisotropic materials [33,34]. Among other post-processing-based imaging methods like phase coherence imaging and spatial compounding imaging, TFM is considered the most robust [35].

1.5 Ultrasonic Coupling and Surface-Adaptive Imaging.

Using an ultrasonic array probe, however, imposes the challenge of ultrasonic coupling between the probe's linear surface and the test object's locally variable curved surface. Along with customizing the probe’s surface [36,37], soft coupling wedges [38], flexible arrays [15,39], or a water jet [16,40], this can be achieved by immersing the test object and probe in water [41]. This way, the test object's irregular surface can be scanned with a wide range of standard ultrasonic array transducers. This comes, however, with another difficulty: The ultrasonic pulse travels through water for a considerable distance before entering the test object. The difference in speeds of sound in water and in the test object, in connection with a non-planar surface, results in considerable diffraction effects. Under knowledge of the test object's surface contour, adjusted travel times can be calculated using Fermat's principle to create a TFM image for the inside of the test object [32,42,43]. This approach provides 2D images using a linear UT array but can also be performed in 3D using a matrix array [44]. In our method, an individual implementation of the adaptive TFM in 2D is used [42]. The 2D method is preferred here since the turbine blade's curvature is usually significantly higher in one direction than it is in the respective perpendicular direction. A square matrix array would, thus, assumably not yield significant advantages compared to the much higher computational effort.

1.6 Automated Evaluation of Images.

Comprehensively scanning the surface instead of testing single points poses an additional challenge: automatic evaluation of the ultrasonic images. For this task, the standard and ever-expanding arsenal of image processing techniques can be applied, but for best results, these techniques must be combined and adapted for a given scenario. Consequently, standard implementations of automatic UT evaluation are limited to using gates and thresholds. In the method presented, we use automated evaluation based on the TFM images initially to provide the outer surface's position and shape. Based on that, images for the interior are calculated, which must be evaluated for identification of the inner surface (back wall). As a result, 2D point clouds for the outer and inner surfaces of the test object are generated for each test position. Consequently, the 2D point clouds are merged in 3D using the known positions and orientations of the test points. For elementary geometries, combining point clouds obtained from 2D images in robot-guided immersion array UT has been shown by Kerr et al. [45]. Here, the selection of points from the images was mainly based on thresholds. The reported errors and standard deviations compared with the reference geometries were in the regime of a few millimeters.

In our case, having more complex surfaces to reconstruct, we can use post-processing on the point clouds based on physical knowledge of the turbine blade or vane, for example, the assumption of a smooth outer surface, to improve the surface reconstruction. From the point clouds, local wall thickness values can be evaluated. This provides a mapping between surface location and wall thickness, which can be well represented using a heat map, which, in the end, provides valuable information for the turbine blade and vane repair.

1.7 Other Nondestructive Testing Methods.

For wall thickness testing, there are a few other NDT methods to consider: Le Bihan et al. showed that turbine blade wall thickness can be evaluated using eddy current testing (ET) [46], reaching a standard deviation (SD) under 25 µm for the quasi-flat areas of the surface and 200 µm for curved areas. However, the method needs to be applied with the probe in contact with the surface and thus has the same drawbacks as the manual UT thickness gauging.

Perhaps the most capable method for volumetric imaging in terms of spatial resolution and reliability is computed tomography (CT) with X-rays. For turbine blades, a method was introduced by Hastie et al. [47] and showed a high accuracy of around 50 µm. The drawback with radiography is that the equipment is usually very costly, and the ionizing radiation involved requires additional safety measures. Using lower energy radiation, for example, terahertz imaging, is not applicable for metals.

Also, wall thickness evaluation is possible using active thermography, as shown by Goldammer et al. [48]. Here, a wall thickness mapping is provided throughout the airfoil, reaching repeatability “better than 0.1 mm”; however, the method is not suitable to extract the outer geometry of the test object.

2 Methodology

2.1 Experimental Setup.

In our experimental setup, a six-axis robotic arm, ABB IRB 1660ID, is used to hold an ultrasonic linear array probe. The robot rests on a custom frame forming a unit with the immersion tank (see Fig. 1). An additional module is attached, providing a working surface and the water storage tank, which provides weight for the stability of the robot and enables the couplant to be reused. Since the probe and test object are immersed in water, a custom probe holder was built to attach to the robot's end effector. This way, the end effector stays out of the water and does not have to be made waterproof. The CAD model of the probe holder and UT probe are used to define the tool center point in the robot software.

Fig. 1
Digital model (left) and photo (right) of testing system
Fig. 1
Digital model (left) and photo (right) of testing system
Close modal

The center frequency of the UT probe, Olympus 7.5L64-64X7-I4-P-7.5-OM, is 7.5 MHz with a relative bandwidth of approximately 67%. The frequency was selected considering a tradeoff: Lower frequency pulses experience less attenuation in the material, while higher frequencies offer better spatial resolution. The array contains a linear arrangement of N = 64 elements with 1 mm pitch and 7 mm elevation. The length of the water path is set to 70 mm since this is roughly the natural focus length of the element's elevation. Given this distance and the transducer's aperture dimensions (64 mm × 7 mm), it becomes clear that a water jet coupling would be much more complicated to realize than the immersion tank. The ultrasonic tests are conducted using a Verasonics Vantage 64 research ultrasound system and the FMC principle. This means that consecutively, each element is used as an ultrasonic transmitter while all elements are used as receivers. That way, A-scans for all possible combinations of ultrasonic elements in the array are acquired, adding up to N2 = 2048 signals. At each test position, one FMC dataset is acquired.

During the scan routine, the ultrasound instrument and the robot are synchronized via electronic trigger signals and a network connection to assign the actual probe position and orientation, which are provided by the robot controller, to each of the FMC data sets. Due to limited buffer size and for easier handling, raw data are stored in segments consisting of the UT data of 100–200 test positions.

For robot path planning, we use the CAD model of the nominal geometry, which provides the advantage that the calculated path can be reused for different parts of the same model. The test object is positioned at predefined attachment points on a base plate in the immersion tank. The base plate itself serves as a work object coordinate system, which is acquired using positioning tips. Furthermore, an optical scan of the base plate with attachment points has been used to compensate for manufacturing errors of the base plate.

The path planning itself is conducted inside ABB's RobotStudio with Machining PowerPac, which allows to create paths that consist of thousands of points while keeping an orientation normal to the test object's surface. While the path planning tool was originally designed for machining applications, it has parameters that allow setting the distance between (a) test points and (b) lines of test points (scan lines). For our UT probe and test object geometry, a point distance of 0.5 mm and a line distance of 4 mm were chosen. Using a linear array probe on a 2-dimensionally curved surface requires including the z-rotation of the tool in the path planning. Here, the array aperture is aligned with the approximate direction of the surface's main curvature. As usual in mechanized UT, the scanning direction of each probe path is set to be perpendicular to the array aperture.

2.2 Post-Processing

2.2.1 Classic Total Focusing Method.

For imaging, a two-step surface-adaptive TFM technique is used in conjunction with automated feature extraction from the images. First, the test object's outer surface must be reconstructed. This is performed based on the classic TFM approach [24]: The intensity I(x,z) at each pixel of the image is a sum of ultrasonic amplitudes evaluated at the individual travel times between the ultrasonic elements and the image pixels:
I(x,z)=|tx,rxhtx,rx((xtxx)2+z2+(xrxx)+z2cl)|
(1)
Here, tx and rx denote the ultrasonic elements used as transmitters and receivers and cl the longitudinal speed of sound. For the x- and z-coordinates, the subscripted symbols stand for the element positions and the non-subscripted for the coordinates of the image pixel. Usually, the transducer surface is considered at z = 0, such that for the element positions, only the x-values appear. In the original TFM equation, the sums are built using the analytical signals:
htx,rx(t)=stx,rx(t)+jH(stx,rx(t))
(2)
which are complex-valued and contain the raw measurement data stx,rx(t) in their real parts and the Hilbert-transformed data H(stx,rx(t)) in their imaginary parts (j is the imaginary unit). This leads to complex-valued image pixels, from which the absolute values are taken, leading to an image with only real positive amplitudes. However, sometimes Hilbert-transforming the raw data is omitted [32] and so is taking the absolute of the pixels [28,49]. In our method, we use the “raw” TFM image built only from the raw ultrasonic signals stx,rx(t):
Iraw(x,z)=tx,rxstx,rx((xtxx)2+z2+(xrxx)+z2cl)
(3)
Furthermore, we build the envelope of the image with the help of the Hilbert-transform in the z-direction, just like the analytical signal in Eq. (2) was built:
Ienv(x,z)=|Iraw(x,z)+jHz(Iraw(x,z))|
(4)

While mathematically it is not obviously equivalent to the original formulation of the TFM, it returns equivalent images and should be justified under the precondition that the z-direction in the images is equivalent to the main direction of wave propagation.

To adapt for element directivity, we use the weighting factors from Le Jeune et al. [42]:
Wjp(θp)=sin(ka2sinθjp)ka2sinθjpcosθjp
(5)
where a is the element width, k=2π/λ the wavenumber, and θp the viewing angle between an element j and the pixel p. The weighing factors, therefore, provide a scaling for each of the summands extracted from the signals:
Iraw(x,z)=tx,rxWtxpWrxpstx,rx((xtxx)2+z2+(xrxx)+z2cl)
(6)

However, calculating images for thousands of test positions can be computationally intensive. Due to the deviations between the test object's actual geometry and the nominal model used for path planning and possible uncertainties in positioning, the distance between the probe and test object surface, though accurately predefined, contains an error. Therefore, for each FMC data set, a quick pre-evaluation is done: The maximum amplitude of the central transmitter–receiver combination is used to determine the “core” depth in which the TFM image is set. Starting from here, an enclosing z-interval is chosen according to the scanned type of surface curvature (convex or concave). This way, the images can be kept as small as possible while still capturing the surface and allowing a sufficient resolution. The images can be restricted laterally as well: Since for a curved surface, the inclination of the surface relative to the probe increases with lateral distance to the array center, reconstruction quality decreases in the peripheral area; therefore, the images can be restricted at the sides to approximately two-thirds of the array aperture.

2.2.2 Surface Extraction.

After calculation of the first stage images, the outer surface's position and shape are extracted. For that, a specialized algorithm has been developed using different image processing techniques. The algorithm takes the TFM images and the pixel's corresponding x- and z-coordinates as input and, for each image, returns a point cloud with surface points identified. The algorithm involves the following steps (see Fig. 2) for each TFM image associated to a test position:

  • A K-means clustering [50] with k = 3 is performed on the image Iraw to identify the regions with the highest probability of containing the surface.

  • A cubic polynomial is fitted to the group of z points with highest associated amplitude for each given x coordinate and which are also contained in the highest mean cluster.

  • The polynomial fit is then evaluated for all the pixel's x-values to provide a local point cloud in the probe coordinate system.

  • Additionally, absolute amplitude values are extracted with the help of the z-direction envelope (Ienv) of the image.

Fig. 2
Surface extraction process for example image: (a) raw TFM image, (b) mask obtained by k-means clustering with three regions, (c) masked TFM image with cubic fit at column maxima (red dotted line), and (d) z-direction envelope of TFM image with amplitudes evaluated at positions of cubic fit (red dotted line)
Fig. 2
Surface extraction process for example image: (a) raw TFM image, (b) mask obtained by k-means clustering with three regions, (c) masked TFM image with cubic fit at column maxima (red dotted line), and (d) z-direction envelope of TFM image with amplitudes evaluated at positions of cubic fit (red dotted line)
Close modal

This extraction procedure is displayed step by step for an example image in Fig. 2.

2.2.3 Adaptive Total Focusing Method.

Once the outer surface has been identified, the adaptive TFM [42] can be applied to calculate images for inside the test object.

First, a reconstruction grid is defined starting at a certain distance behind the outer surface, spanning the area of the expected occurrence of the test object's inner surface (back wall indication). Laterally, the image is refined to the width of the identified surface minus a defined lateral amount, which has the effect of preventing artifacts where no comprehensive outer surface has been identified. For the calculation of the adapted travel times, Fermat's principle is used: In the main brackets of the following equation, the travel time from an ultrasonic element e with the coordinates (xe,0) to the pixel P with coordinates (x,z) passing through a point Sj on the surface with coordinates (xSj,zSj) with sound velocities c1 above and c2 below the surface is calculated. From that, the shortest travel time with respect to changing the surface through-point is selected:
teP(xP,zP)=minSj((xSjxe)2+zSj2c1+(xPxSj)2+(zPzSj)2c2)
(7)
A visual aid to this principle is Fig. 3. These travel times are calculated for all paths from elements to pixels of the sub-surface image. To obtain travel times from transmitting element to pixel to receiving element, these can be linearly combined under the assumption of point reflectors, forming all “round trips” possible:
ttx,rxP=ttxP+trxP
(8)
Fig. 3
Schematic diagram for the calculation of adapted travel times between ultrasonic elements and sub-surface image pixels using Fermat's principle for one combination of transmitting (tx) and receiving (rx) element and a pixel P (recreated from Ref. [42])
Fig. 3
Schematic diagram for the calculation of adapted travel times between ultrasonic elements and sub-surface image pixels using Fermat's principle for one combination of transmitting (tx) and receiving (rx) element and a pixel P (recreated from Ref. [42])
Close modal

One of these round trips is shown in Fig. 3. The dashed lines mark the travel times also calculated but discarded by Fermat's principle. With all the travel times calculated, they can be used as an argument in Eq. (6). For the weighting factors, a simplification is made to avoid having to propagate all the beam cones through the surface: Here, for the angles θjP, the viewing angles from the element j to the selected through-point Sj was taken instead of the direct viewing angle from j to the pixel P.

2.2.4 Back Wall Extraction and Global Filtering.

On the second set of TFM images, once again, a surface identification is performed, this time for the back wall. Since the outer and inner surfaces have different characteristics, a different algorithm must be used. The back wall identification algorithm involves the following steps:

  • A K-means clustering (with k = 4) is used on the TFM image's z-direction envelope (Fig. 4(a)) to identify the regions with the highest probability of containing a back wall surface (Fig. 4(b)).

  • A Gaussian process regression (GPR) is performed to fit a piecewise Gaussian curve to the z-coordinate of the minimum amplitude point for each column of the TFM image (Fig. 4(c)). Choosing the minimum instead of the maximum amplitude is motivated by the back wall echo always being inverted in phase.

  • Finally, the GPR curve is masked with the highest mean cluster selected in the first step to return a back wall point cloud (Fig. 4(d)). To avoid extracting outliers, points whose associated absolute amplitude drops below a percentage of the moving average across x coordinates are also filtered out.

Fig. 4
Backwall extraction process for example image: (a) z-direction envelope of second-stage TFM image, (b) mask obtained by k-means clustering of (a) with four regions, (c) raw second-stage TFM image with Gaussian process regression fit (red dotted line) for column minima, and (d) image and fit from (c), masked with highest region from (b). The red dots in this image are the extracted back wall points.
Fig. 4
Backwall extraction process for example image: (a) z-direction envelope of second-stage TFM image, (b) mask obtained by k-means clustering of (a) with four regions, (c) raw second-stage TFM image with Gaussian process regression fit (red dotted line) for column minima, and (d) image and fit from (c), masked with highest region from (b). The red dots in this image are the extracted back wall points.
Close modal

After identification of the back wall points in all images, they can then be filtered from the global perspective. Two filters were established:

  1. The back wall images are compared by their absolute maximum values (one value per image). These maxima are sorted by magnitude and plotted over a number of images (solid line, Fig. 5 (top)). A linear fit is performed (dotted line). A condition is established: Values lower than 90% of the zero-order term of the fit are filtered out (dashed horizontal line), thus eliminating the respective images from the evaluation.

  2. Furthermore, the spread of the back wall points is considered (Fig. 5 (bottom)): While the back wall in each image is not expected to be continuous, it is indeed expected not to be scattered. An indicator for that is the SD of the z-coordinates of the extracted back wall points of each image. After calculating these SD for each image (again resulting in one value per image), a global average SD value can be determined and used as a threshold. The condition determined is that the individual SD from each image must not be greater than 1.7 times the global SD for an image to be selected. This threshold is marked by the dotted horizontal line in Fig. 5(b).

Fig. 5
Result of global filtering for back wall images of scan 1. Top: Thresholding by sorting the maximum amplitudes for each image (solid line) and linear fit (dotted line). The dashed horizontal line indicates the resulting amplitude threshold (see text). Bottom: Standard deviation (scattered dots) of extracted back wall z-coordinates for each image with threshold (dotted horizontal line). Each vertical dashed line represents the start of a new scan line.
Fig. 5
Result of global filtering for back wall images of scan 1. Top: Thresholding by sorting the maximum amplitudes for each image (solid line) and linear fit (dotted line). The dashed horizontal line indicates the resulting amplitude threshold (see text). Bottom: Standard deviation (scattered dots) of extracted back wall z-coordinates for each image with threshold (dotted horizontal line). Each vertical dashed line represents the start of a new scan line.
Close modal

These two conditions help avoid artifacts in the final point cloud, especially when there is no back wall echo present in an image. A precondition for that is that the images with sufficient feature quality are dominant in the scan. The percentage values for the thresholds have been found empirically.

2.2.5 Merging and Adjustment of Point Clouds.

With the images calculated and the surfaces extracted from the individual images, all data acquired are merged in 3D space. As the surface point clouds are still in the local probe coordinate system, they first undergo a transformation into the robot's work object coordinate system, considering the probe position and orientation of each test point. A second transformation is global for all points and changes the frame of reference to that of the initial CAD model. After these transformations, the 3D outer surface point cloud undergoes an adjustment, turning it into a surface, which consists of z values over a regular x-y-grid. This can (a) smooth out some fluctuation in the reconstructed surface, which is explained in the following, and (b) enable the definition of surface normal vectors, which are needed for the later evaluation of wall thickness. The mentioned fluctuation is due to an oversampling effect mainly in the regions with lower surface curvature: In the images, the width of the region where surface points can be reliably identified depends inversely on the surface curvature (higher curvature means smaller effective aperture). Since the distance between scan lines was set small enough to avoid gaps in the surface reconstruction, regions with lower surface curvature are thus reconstructed by multiple adjacent scan lines. This introduces noise caused by positioning and alignment errors. To address this, we resample the point cloud on a regular x-y-grid. In each grid cell, a weighted average of the z-position of all points in that cell is calculated, from which the partial surface’s normal vectors are determined. In a second iteration, another, more precise search is conducted, selecting and averaging all points inside the unit cell perpendicular to the local surface normal. The averaging is justified under the assumption of a smooth surface with respect to the grid resolution.

After the final point clouds for the inner and outer surfaces are provided, the wall thickness values are calculated locally. For each point in the outer surface point cloud, a partner point in the inner surface point cloud is searched along the normal direction of the outer surface point. This is restricted to a cone with an opening angle of 10 deg. When the closest point is found, the distance projected along the surface normal is logged as wall thickness corresponding to the outer surface point. For comparison with the reference surface data (see below), the backwall points are repositioned to be on their respective partner point's surface normal. This procedure is motivated by the following: The testing application demands surface points and corresponding wall thickness (WT) values to be provided. However, since back wall surfaces are required for thickness evaluation, it is natural to identify the corresponding perpendicular back wall locations for greater insight into the derived thickness evaluations. Surface points where no back wall can be identified are kept as surface points and will get a NaN value as local wall thickness. Back wall points that have not been selected as a “partner point” to a surface point are here discarded. The complete post-processing is schematically depicted in Fig. 6.

Fig. 6
Schematic diagram for the complete post-processing of the ultrasonic testing data
Fig. 6
Schematic diagram for the complete post-processing of the ultrasonic testing data
Close modal

2.3 Specimens and Scan Series.

A special aim of the project was to develop the proposed method to work on real industry parts. Therefore, two turbine vanes of a modern gas turbine have been provided. The material of the vanes is a Ni-based superalloy and is conventionally cast, such that the crystal grain structure is equiaxed, which means that the orientations of the grains are randomly distributed. The acoustic properties of the material can thus be considered isotropic. The airfoil of the specimens has been opened partially in order to acquire geometric reference data (see next paragraph). The areas of interest of one of the two specimens are displayed in Fig. 7. As usual, in this stage of the MRO process, the thermal barrier coating (TBC) has been removed from the vanes.

Fig. 7
One of the two tested specimens. Left: convex side, right: concave side. The airfoil has partially been opened to acquire wall thickness reference data using an optical scan. The rulers in the images have a length of 10 mm.
Fig. 7
One of the two tested specimens. Left: convex side, right: concave side. The airfoil has partially been opened to acquire wall thickness reference data using an optical scan. The rulers in the images have a length of 10 mm.
Close modal

The desired proximity to the industrial application, however, poses two challenges: First, the real part includes diagonal bars on the inner surface, which have the function of introducing turbulence in the cooling gas inside the vane's cavity. These are challenging to reconstruct with UT due to their size and geometry. Second, each manufactured turbine blade or vane will deviate from the nominal geometry provided by the CAD model.

On the two specimens, a series of 20 scans were performed (see Table 1). This includes 10 scans for each vane: five on the referenced zones on the convex side and five on the concave side, respectively. During these five scans for each side, which will be named “5-scan subseries” from here, the vane was newly repositioned on the base plate for the first three scans and then scanned again twice with the same positioning.

Table 1

Overview of scans conducted on two specimens

Scan number1234567891011121314151617181920
SpecimenVane 1Vane 2Vane 1Vane 2
Test zoneConvex sideConvex sideConcave sideConcave side
Repositionedyyynnyyynnyyynnyyynn
Scan number1234567891011121314151617181920
SpecimenVane 1Vane 2Vane 1Vane 2
Test zoneConvex sideConvex sideConcave sideConcave side
Repositionedyyynnyyynnyyynnyyynn

2.4 Reference Data.

To evaluate the accuracy of the proposed method, reference data from optical scanning have been prepared. In specific areas of the specimen, the outer surface has been removed such that the inner surface could be scanned (see Fig. 7). The scans were conducted using a Blue Light Sensor providing the inner surface data for an area on the convex side and another area on the concave side. The point clouds created with the UT method were compared with the reference data in the following way:

  1. The outer and inner surface point clouds (UT) are polygonised. To avoid polygonization distorting the point clouds, the polygonization parameters are chosen such that only the existing points of the UT point clouds are connected to polygons, and no additional points are created. After polygonization, the outer and inner surface meshes are merged into one data set.

  2. The resulting mesh dataset is aligned to the optical scan by means of a best fit to be able to evaluate the specimen's position deviation separately from its shape deviation.

  3. The position-related deviation calculated by the best fit is reported as translational deviation x, y, z and rotational deviation Phi, Theta, Psi.

  4. For the optical scan mesh and the UT mesh, the local wall thickness values are calculated and compared with each other. The local deviation of the UT wall thicknesses from optical scan wall thicknesses is presented as a color plot (see Fig. 12). In addition, the mean value and standard deviation of all values are calculated and reported for each scan. These values are indicative of the accuracy of the respective scan compared to the optical scan. Furthermore, for each color plot, the distribution of the values is displayed in the form of a histogram.

Fig. 12
Deviation of UT wall thickness values from reference wall thickness values for scans (a) 1, (b) 6, (c) 11, and (d) 16. The green areas indicate an agreement of UT and reference data of ±0.1 mm. The color maps span an interval of −0.4 mm to 0.4 mm.
Fig. 12
Deviation of UT wall thickness values from reference wall thickness values for scans (a) 1, (b) 6, (c) 11, and (d) 16. The green areas indicate an agreement of UT and reference data of ±0.1 mm. The color maps span an interval of −0.4 mm to 0.4 mm.
Close modal

3 Experimental Results

3.1 Visual Representation of Back Wall Reconstruction.

Before evaluating the reconstruction quantitatively, let us have a look at a possible representation of a reconstructed geometry, as depicted in Fig. 8. Here, a colored point is drawn on the outer surface grid (transparent mesh) wherever a wall thickness could be extracted under the filters discussed. The corresponding color shows the absolute wall thickness values in millimeters. The blank spaces parallel to the x-axis at y = 190 mm and y = 210 mm provide no wall thickness since, here, the material bridges from one outer surface to the other. Due to the geometry, the sound path is too narrow to receive a measurable echo here. Furthermore, the reconstructed inner surface is interrupted by diagonal bars—the aforementioned turbulators. As expected, their geometry at the inner surface cannot be reconstructed as reliably as the parallel parts of the back wall. However, through the absence of the signal, the turbulator's position and shape can be identified visually.

Fig. 8
Representation of wall thickness mapping. The gray grid represents the reconstructed outer surface. Colored points represent reconstructed back wall points where the wall thickness could be extracted. The color code shows the local wall thickness in millimeters.
Fig. 8
Representation of wall thickness mapping. The gray grid represents the reconstructed outer surface. Colored points represent reconstructed back wall points where the wall thickness could be extracted. The color code shows the local wall thickness in millimeters.
Close modal

3.2 Comparison With Reference Data

3.2.1 Position-Related Deviation.

Comparing the results with the reference data, let us first discuss the position-related deviation before addressing the accuracy in wall thickness evaluation. Figures 9 and 10 display the transitional and rotational transformation parameters that were identified during the 6-degree-of-freedom (6DOF) alignment. The transitional values are in the sub-millimeter range and the rotational values in the sub-degree range. It should be mentioned that the reference data itself have been aligned to the CAD model using the entire surface, minimizing the total quadratic error. Since, in the experiment, the specimen was laid down on either one of two sides, a perfect match between the coordinate systems was initially not expected. However, the distribution of the alignment parameters over the subsequent scans can show how well the positioning can be reproduced. Comparing these values within each 5-scan subseries (i.e., 1–5, 6–10, 11–15, and 16–20), no obvious difference can be seen between the first three and the fourth and fifth scan of each subseries. The parameters are either relatively stable for all five scans of a subseries (e.g., scans 6–10), or there are a few parameters that are more variable within the subseries (e.g., scans 11–15). Remembering that for the second and third scans of each subseries the test object was repositioned, while for the fourth and fifth scans it was not, this indicates that an influence of the repositioning cannot be seen. Therefore, it appears that judging by this method, the repositioning tolerance is smaller than the cumulative tolerance of the evaluation method. Comparing the rotational values yields the same results (Fig. 10). Furthermore, it is observed that for the translational parameters, the highest values can be found in the z-coordinate in the second subseries (scans 6–10) and in y in the third subseries (scans 11–15). While for scans 6–10, the values are stable around x = 0.9 mm, there is more spread in the parameters for scans 11–15, containing the absolute highest value with y = 1.5 mm for scan number 12. Comparing that with the rotational values, there is an obvious accumulation of the absolute highest angle values for θ = −0.5 deg in scans 6–10, while for scans 11–15, all angles are spread between 0 deg and 0.2 deg. The exact 6DOF alignment is sensitive to many parameters since it returns the positions and orientations based on minimizing the average square error between all points of the clouds, which makes it difficult to interpret single parameters separately. However, the subseries 2 and 3 (respectively 1 and 4) should have the least in common since there is a change in the tested side and specimen. Therefore, it appears reasonable that the highest differences in positioning parameters exist between these subseries. In general, the alignment parameters differ more when compared between the four subseries, which will be addressed further in the discussion section.

Fig. 9
Translational parameters from 6-deg-of-freedom alignment for the 20 scans. The dotted vertical lines separate the different 5-scan subseries.
Fig. 9
Translational parameters from 6-deg-of-freedom alignment for the 20 scans. The dotted vertical lines separate the different 5-scan subseries.
Close modal
Fig. 10
Rotational parameters from 6-deg-of-freedom alignment
Fig. 10
Rotational parameters from 6-deg-of-freedom alignment
Close modal

3.2.2 Wall Thickness Deviation.

Finally, the quality of the wall thickness values is evaluated. As stated above, wall thickness values were extracted from both the UT reconstructed surfaces and reference surfaces. The meshes created from the UT point clouds of scans 1, 6, 11, and 16 are depicted in Fig. 11. The color plot shows the absolute local wall thickness values with their distribution in the histogram on the right. Here, the wall thickness values were determined, as described in step 4 of Sec. 2.4, to be able to compare them to the reference wall thickness values. Especially in Figs. 11(a) and 11(b), there are gradients in wall thickness values visible in the y-direction, similar to those in Fig. 8. These gradients appear to show true changes in wall thickness, which is to be evaluated in the reference data comparison. In Fig. 11(c), the wall thickness values seem to change in certain “steps” between the reconstructed areas (from left to right), which is less the case in Fig. 11(d). Also, this phenomenon is to be evaluated in the following. Differences in wall thickness between the two specimens are also visible, as well as regions where the reconstruction works only for one of the two specimens (compare right-hand sides of Fig. 11(a) / 11(c) and Fig. 11(b) / 11(d), respectively). However, most of the zones of interest have similar reconstruction qualities for both specimens.

Fig. 11
Evaluated wall thickness meshes from UT scans (a) 1, (b) 6, (c) 11, and (d) 16 as prepared for the comparison with optical reference data. The color codes show the determined wall thickness values from 0 mm to 4 mm.
Fig. 11
Evaluated wall thickness meshes from UT scans (a) 1, (b) 6, (c) 11, and (d) 16 as prepared for the comparison with optical reference data. The color codes show the determined wall thickness values from 0 mm to 4 mm.
Close modal

The wall thickness mapping is then locally compared to the mapping from the reference data, as shown in Fig. 12. These diagrams show the local deviation between the wall thickness values from the UT mesh with those from the reference mesh. Values inside a span of ±0.1 mm around zero are represented in green color. Most values lie within −0.2 mm and +0.1 mm; however, there are some outliers. The fact that the compared surface appears to be higher in resolution than the UT scans themselves is due to the higher spatial resolution of the reference data. The aforementioned gradients in WT of Figs. 11(a) and 11(b) as well as the “steps” in WT in Figs. 11(c) and 11(d) are not visible in the respective diagrams in Fig. 12. Since this figure shows the deviations between UT wall thickness and reference wall thickness, this indicates at least qualitatively that these phenomena are real features in the actual geometry of the specimens.

The local values of wall thickness deviation can furthermore be compacted into one average and one variance measure for each of the 20 scans. These numeric values are depicted in Fig. 13, where the mean deviation is depicted by the black squares and their statistic spread by the error bars (± one standard deviation). The mean difference reaches from −0.05 mm to −0.1 mm and can be interpreted as a systematic error. The standard deviation reaches from 0.06 mm to 0.16 mm. Both mean and standard deviation stay relatively constant within each 5-scan subseries. Given this low spread, it makes sense to average these values again to represent the reconstruction quality for each of the subseries, as depicted in Table 2. The spread of scans 6–10 with a mean value of 0.16 mm is significantly higher than the other values. Here, Fig. 12(b) reveals high wall thickness deviations (dark blue lines) in the areas with the smallest turbulators. These lead to an increase in the standard deviation. The root cause of these higher deviations is that for the given datasets, the position of the gaps in UT point clouds, which represent turbulators, is significantly deviating from turbulator positions acquired by optical scans. The evaluation algorithm calculates and compares wall thickness values at each location. At the given locations, the fact that the optical scan does contain turbulators while the UT reconstruction does not introduces higher deviation displayed for such mismatching areas. Potential sources for this behavior are overall inaccuracies in the UT dataset, possibly a distortion in form, and the 6DOF best-fit alignment, which has insufficient information along surfaces with low curvature.

Fig. 13
Average deviation between UT reconstructed and reference wall thickness values. The error bars show ±one standard deviation.
Fig. 13
Average deviation between UT reconstructed and reference wall thickness values. The error bars show ±one standard deviation.
Close modal
Table 2

Averaged statistical results from the scan series

Number of scan1234567891011121314151617181920
SpecimenVane 1Vane 2Vane 1Vane 2
Test zoneConvex sideConvex sideConcave sideConcave side
Mean wall thickness deviation (mm)−0.05−0.06−0.06−0.10
Mean spread (mm) 0.06 0.16 0.09 0.09
Number of scan1234567891011121314151617181920
SpecimenVane 1Vane 2Vane 1Vane 2
Test zoneConvex sideConvex sideConcave sideConcave side
Mean wall thickness deviation (mm)−0.05−0.06−0.06−0.10
Mean spread (mm) 0.06 0.16 0.09 0.09

3.3 Scanning and Computation Times.

For a potential industry application, efficiency is important; therefore, we state and discuss the scanning and computation times of our proposed method. Obviously, these durations are highly dependent on the selected testing and reconstruction parameters as well as the used hardware. In our case, scanning 1886 probe positions on the convex side took approximately 4.5 min and 1114 positions on the concave side 3 min. Given the share of the scanned zones on each side was approximately 15%, the scanning time for the airfoil of the used vanes extrapolates to approximately 50 min. This is sufficient for laboratory scale but should be optimized when applied in industry, where many parts must be scanned consecutively. In our setup, the main bottleneck in scanning speed is probably point-by-point scanning, stopping at each point instead of scanning continuously and measuring during movement. The scanning speed could probably be improved by moving the probe continuously, obtaining UT data along the way. This would be possible using a different synchronization concept; however, small breaks in the scan routine every few hundred points would still be required to store the data on the hard drive.

Regarding reconstruction time, it is widely known that using the TFM is computationally intensive. There are devices present on the market that can provide live TFM imaging (about 30 frames per second) under certain circumstances. Our method, however, uses an even more expensive two-stage TFM and was implemented and run on a computer with an Intel(R) Core(TM) i7-6700 CPU @ 3.40 GHz with 16 GB of memory. The calculations of the surface and back wall images were parallelized with the matlab Parallel Computing toolbox to use the four CPU cores in parallel. The total processing time for the two scanned zones was 108 min and 48 min, respectively. Extrapolated to a dataset of the whole airfoil of a blade, comprising around 20,000 test positions, the theoretical computation time on our computer would be approximately 17.4 h. Here, the potential for acceleration depends on the hardware used. Using powerful hardware, e.g., cloud computing, it should be possible to reduce the computation time to be in the same regime as the scanning time. It is also possible to run the scanning and computation independently from each other since handling the test objects requires human interaction while running the algorithms does not.

4 Discussion

Having compared the results of the 20 scans of two turbine vanes with geometric reference data, the position-related deviation and reconstructed wall thickness deviation have been evaluated. The position-related deviation showed relatively low variance for repeated scans of the same test zone and specimen. Within each subseries, there is no clear indication of an influence of repositioning the specimen. Changing the test zone or specimen, however, led to different positioning parameters. With respect to the test zone, this is probably because to test the other side, the specimen is turned around and different attachment points are used. However, the variation in parameters with the change of specimen comes with the uniqueness of every individual part. In total, the misalignment error is mostly influenced by uncertainties of attachment, the actual geometry of the specimen, robot positioning tolerance, ultrasound measurement error, polygonization, and best-fit alignment. Apparently, by an evaluation of the reconstructed surfaces alone, these influences cannot be isolated from each other. The presented alignment parameters are, therefore, rather an indicator for the overall repeatability tolerance of the system, which showed to be in the sub-millimeter and sub-degree range for most scans. Comparing the variations inside the subseries with the ones between subseries, it appears that the “noise” from reproducing the same scan, even with the specimen repositioned, is about a magnitude smaller than the one induced by a change of the test zone or specimen. In conclusion, that shows that the particular geometry deviation of each test object has a measurable influence on reconstruction quality. This is probably an interplay of the overall shape deviation and local shape deviation at the attachment surfaces. It is possible that they can either cancel out or add up for each individual test object. For increased accuracy, an improved universal attachment or individual path planning based on the actual geometry could be developed. In conclusion, it appears that for the proposed method to work, scan path planning using only the nominal model of the test object is sufficient. However, in further works, it should be analyzed if and how individual path planning based on the actual geometry of the test object can improve reconstruction quality and increase the precision and reliability of the wall thickness evaluation.

For reconstructed wall thickness deviation, the comparisons were compacted to a mean deviation and mean deviation spread (standard deviation) for each 5-scan subseries since the values inside the subseries were relatively stable. Overall, the concave side appears to be more challenging to reconstruct since it shows a higher spread (neglecting the discussed 0.16 mm outlier for the convex side on vane 2). This is probably because it offers less contiguous areas and more complex features. Also, the test zone on the concave side is slightly longer and has an increased range of wall thickness values to possibly reconstruct, allowing more room for error (compare Figs. 11(a) and 11(b) with Figs. 11(c) and 11(d)).

Over all scans, there is a mean wall thickness deviation from reference data, ranging from −0.05 mm to −0.1 mm, which can be identified as a systematic error. The most probable cause of this error is a slightly different actual value of the speed of sound in the material. It is known that for the material used, the speed of sound can deviate locally by a small percentage and is only isotropic on a scale above the grain size. The systematic error shows that the mean speed of sound c = 5870 m/s, which was used in the calculations, was apparently not optimal. For future works, an adjustment of the value by the discussed systematic error could be made, or the speed of sound could be measured more precisely beforehand with a suitable method.

Compared to similar NDT methods that provide a local wall thickness mapping, our method is in the same regime as the reported CT [47] and thermography [48] methods. Modern automated single-point testing ultrasound methods can reportedly be even more accurate [17] since it does local adjustments on the test points but does not provide comprehensive surface information.

4.1 Limitations and Outlook.

In our experimental setup, the positioning of the specimen has taken advantage of the fact that this type of turbine vane has plane faces, which allows it to be laid on either of two sides stably. Consequently, the positioning could be realized simply by using a plate with a few bolts attached as stops. This comes with the drawback of having to turn the specimen between testing of the concave and convex sides and, therefore, having to align and merge two datasets. Here, the simplicity of the setup was preferred. However, in further works, the possibility of a different solution should be explored, in which the test object would be vertically mounted, being scanned by a horizontally held probe in a circular motion. That way, the test zone could be scanned in one run, and the results would naturally be in one common coordinate system.

Another challenging aspect was the evaluation of the wall thickness data. Using real turbine vanes for evaluation, the comparison to the reference data was reliant on the evaluation software of the optical scanner and took additional steps to process the UT results. Due to the complex nature of the 3D shape comparison of discrete data, the evaluation had to be condensed down to two statistical values for each scan, which is comprehensible but also limited information. Looking forward, an evaluation of isolated cases of the distinct geometry features could possibly provide more information about the accuracy of the system.

Having introduced a demonstrator testing system raises the question about industrial applicability. While the system was developed with industrial application in mind, some aspects have been intentionally kept at the laboratory scale. Handling and clamping of the test objects were kept simple and would have to be developed by the industrial user individually, depending on the tested parts. Regarding the specially prepared specimens, the surface openings were only made to provide reference wall thickness values to evaluate the proposed method and are not necessary for industrial application. However, on all materials to be tested, a calibration should be made to acquire the actual speed of sound. For turbine blade materials with anisotropic acoustic properties, more complex adjustments have to be made to the imaging algorithms, provided the crystal orientation is known. Ultrasonic imaging for anisotropic materials is an ongoing research topic, as mentioned in the literature review. Furthermore, the consequences of immersing the test objects in water have not been investigated. Since the components are tested without the TBC and using anti-corrosion additives, thoroughly drying the components should be sufficient to prevent water-induced damage; however, further investigations should be made in the individual case. Finally, the economic value of implementing the testing system in the industrial MRO process can only be evaluated by the implementing user.

5 Conclusion

In this article, we have shown how an automated ultrasonic NDT system for the evaluation of wall thickness for complex-shaped parts can be realized. In summary, the possibilities and flexibility of a robot arm were combined with modern ultrasound imaging and data processing methods to achieve a locally resolved mapping of the test object's wall thickness. Compared to the manual, pointwise method frequently used in turbine blade and vane thickness evaluation, this is a big step toward higher accuracy, reproducibility, and reliability and goes hand in hand with NDT 4.0, where comprehensive, digitalized testing data are the basis of modern maintenance processes. On two turbine vanes, a total of 20 scans were performed, five for the convex and concave sides each. From these scans, outer and inner geometries have been reconstructed and the wall thickness values determined locally. The reconstructed results have been compared to reference data in alignment and accuracy of determined wall thickness. The constructed wall thickness values deviated from the reference wall thicknesses between 0.05 mm and 0.1 mm on average, with a standard deviation of 0.06 mm to 0.16 mm.

Acknowledgment

The authors would like to thank the project partners Siemens Energy Global GmbH & Co. KG, KleRo Roboterautomation GmbH, and Datalyze Solutions GmbH for their cooperation and contribution as well as the Werner von Siemens Centre for Industry and Science e.V. (WvSC) for the organization and patronage.

Funding Data

  • The project MRO2.0 is partially financed by the ProFIT fund of Investitionsbank Berlin (IBB).

  • Co-funded by the European Regional Development Fund (ERDF).

Conflict of Interest

There are no conflicts of interest.

Data Availability Statement

The datasets generated and supporting the findings of this article are obtainable from the corresponding author upon reasonable request.

References

1.
Chowdhury
,
T. S.
,
Mohsin
,
F. T.
,
Tonni
,
M. M.
,
Mita
,
M. N. H.
, and
Ehsan
,
M. M.
,
2023
, “
A Critical Review on Gas Turbine Cooling Performance and Failure Analysis of Turbine Blades
,”
Int. J. Thermofluids
,
18
(
1
), p.
100329
.
2.
Rao
,
N.
,
Kumar
,
N.
,
Prasad
,
B.
,
Madhulata
,
N.
, and
Gurajarapu
,
N.
,
2014
, “
Failure Mechanisms in Turbine Blades of a Gas Turbine Engine—An Overview
,”
Int. J. Eng. Res. Dev.
,
10
(
8
), pp.
48
57
.
3.
Burghardt
,
A.
,
Kurc
,
K.
,
Szybicki
,
D.
,
Muszyńska
,
M.
, and
Szczęch
,
T.
,
2017
, “
Robot-Operated Inspection of Aircraft Engine Turbine Rotor Guide Vane Segment Geometry
,”
Tech. Gaz.
,
24
(
2
), pp.
345
348
.
4.
Weber
,
W. H.
,
Mair
,
H. D.
,
Jansen
,
D.
, and
Lombardi
,
L.
,
2013
, “
Advances in Inspection Automation
,”
AIP Conf. Proc
,
1511
(
1
).
5.
Sattar
,
T.
,
2010
, “
Robotic non-Destructive Testing
,”
Ind. Rob.
,
37
(
5
), p.
420
.
6.
Sattar
,
T. P.
, and
Brenner
,
A. A.
,
2009
, “
Robotic System for Inspection of Test Objects With Unknown Geometry Using NDT Methods
,”
Ind. Rob.
,
36
(
4
), pp.
340
343
.
7.
Morozov
,
M.
,
Mineo
,
C.
,
Pierce
,
S. G.
,
Nicholson
,
P. I.
, and
Cooper
,
I.
,
2013
, “
Computer-Aided Tool Path Generation for Robotic Non-Destructive Inspection
,”
Conf. Br. Inst. Non-Destructive Test
,
Telford, UK
,
Sept. 10–12
.
8.
Mineo
,
C.
,
Pierce
,
S. G.
,
Wright
,
B.
,
Nicholson
,
P. I.
, and
Cooper
,
I.
,
2015
, “
Robotic Path Planning for Non-Destructive Testing of Complex Shaped Surfaces
,”
AIP Conf. Proc.
,
1650
(
1
), pp.
1977
1987
.
9.
Mineo
,
C.
,
Pierce
,
S. G.
,
Nicholson
,
P. I.
, and
Cooper
,
I.
,
2016
, “
Robotic Path Planning for Non-Destructive Testing—A Custom MATLAB Toolbox Approach
,”
Rob. Comput. Integr. Manuf.
,
37
(
1
), pp.
1
12
.
10.
Mineo
,
C.
,
Pierce
,
S. G.
,
Nicholson
,
P. I.
, and
Cooper
,
I.
,
2017
, “
Introducing a Novel Mesh Following Technique for Approximation-Free Robotic Tool Path Trajectories
,”
J. Comput. Des. Eng.
,
4
(
3
), pp.
192
202
.
11.
Seguin-Charbonneau
,
L.
,
Walter
,
J.
,
Dubois
,
P. O.
,
St-Jacques
,
D.
, and
Brassard
,
M.
,
2021
, “
Automated Path Generation for Robotic UT Inspection of CFRP Components
,”
12th International Symposium on NDT in Aerospace
,
Williamsburg, VA
,
Oct. 6–8
.
12.
Brassard
,
M.
,
Séguin-Charbonneau
,
L.
,
Walter
,
J.
, and
Cormier
,
G.
,
2023
, “
Automatic Methods for Ultrasonic Scanning Paths Generation
,”
Proceedings of the European Conference on Non-Destructive Testing (ECNDT)
,
1
(
1
).
13.
Riise
,
J.
,
Mineo
,
C.
,
Pierce
,
S. G.
,
Nicholson
,
P. I.
, and
Cooper
,
I.
,
2019
, “
Adapting Robot Paths for Automated NDT of Complex Structures Using Ultrasonic Alignment
,”
AIP Conf. Proc.
,
2102
(
1
), p.
040006
.
14.
Poole
,
A.
,
Sutcliffe
,
M.
,
Pierce
,
G.
, and
Gachagan
,
A.
,
2022
, “
Autonomous, Digital-Twin Free Path Planning and Deployment for Robotic NDT: Introducing LPAS: Locate, Plan, Approach, Scan Using Low Cost Vision Sensors
,”
Appl. Sci.
,
12
(
10
), p.
5288
.
15.
Vasilev
,
M.
,
MacLeod
,
C. N.
,
Loukas
,
C.
,
Javadi
,
Y.
,
Vithanage
,
R. K. W.
,
Lines
,
D.
,
Mohseni
,
E.
,
Pierce
,
S. G.
, and
Gachagan
,
A.
,
2021
, “
Sensor-Enabled Multi-Robot System for Automated Welding and In-Process Ultrasonic NDE
,”
Sensors
,
21
(
15
), p.
5077
.
16.
Mineo
,
C.
,
Pierce
,
S. G.
,
Wright
,
B.
,
Cooper
,
I
, and
Nicholson
,
P. I.
,
2015
, “
PAUT Inspection of Complex-Shaped Composite Materials Through Six DOFs Robotic Manipulators
,”
Insight-Non-Destr. Test. Cond. Monit.
,
57
(
3
), pp.
161
166
.
17.
Zhou
,
B.
,
Tian
,
T. T.
,
Zhu
,
G.
,
Zhao
,
J. B.
, and
Liu
,
D. H.
,
2022
, “
An Ultrasonic Testing Method for Wall Thickness of Turbine Blades
,”
Measurement
,
198
(
1
), p.
111357
.
18.
Frederick
,
J. R.
,
Fairchild
,
R.
, and
Anderson
,
B.
,
1977
, “
Improved Ultrasonic Nondestructive Testing of Pressure Vessels. Annual Progress Report, August 1, 1975–July 31, 1976
,” Michigan Univ., Ann Arbor (USA). Dept. of Mechanical Engineering.
19.
Schmitz
,
V.
,
Chakhlov
,
S.
, and
Müller
,
W.
,
2000
, “
Experiences With Synthetic Aperture Focusing Technique in the Field
,”
Ultrasonics
,
38
(
1
), pp.
731
738
.
20.
Macovski
,
A.
,
1979
, “
Ultrasonic Imaging Using Arrays
,”
Proc. IEEE
,
67
(
4
), pp.
484
495
.
21.
Defranould
,
P.
, and
Souquet
,
J.
,
1979
, “Ultrasonic Array Design and Performance,”
Echocardiology
,
C. T.
Lancée
, ed.,
Springer
,
Dordrecht, Netherlands
, pp.
395
412
.
22.
Clay
,
A. C.
,
Wooh
,
S.-C.
,
Azar
,
L.
, and
Wang
,
J.-Y.
,
1999
, “
Experimental Study of Phased Array Beam Steering Characteristics
,”
J. Nondestr. Eval.
,
18
(
2
), pp.
59
71
.
23.
Oralkan
,
O.
,
Ergun
,
A. S.
,
Johnson
,
J. A.
,
Karaman
,
M.
,
Demirci
,
U.
,
Kaviani
,
K.
,
Lee
,
T. H.
, and
Khuri-Yakub
,
B. T.
,
2002
, “
Capacitive Micromachined Ultrasonic Transducers: Next-Generation Arrays for Acoustic Imaging?
,”
IEEE Trans. Ultrason. Ferroelectr. Freq. Control
,
49
(
11
), pp.
1596
1610
.
24.
Holmes
,
C.
,
Drinkwater
,
B.
, and
Wilcox
,
P.
,
2004
, “
The Post-Processing of Ultrasonic Array Data Using the Total Focusing Method
,”
Insight-Non-Destr. Test. Cond. Monit.
,
46
(
11
), pp.
677
680
.
25.
Holmes
,
C.
,
Drinkwater
,
B. W.
, and
Wilcox
,
P. D.
,
2005
, “
Post-Processing of the Full Matrix of Ultrasonic Transmit–Receive Array Data for Non-Destructive Evaluation
,”
NDT&E Int.
,
38
(
8
), pp.
701
711
.
26.
Sutcliffe
,
M.
,
Weston
,
M.
,
Dutton
,
B.
,
Charlton
,
P.
, and
Donne
,
K.
,
2012
, “
Real-Time Full Matrix Capture for Ultrasonic Non-Destructive Testing With Acceleration of Post-Processing Through Graphic Hardware
,”
NDT&E Int.
,
51
(
1
), pp.
16
23
.
27.
Bannouf
,
S.
,
Robert
,
S.
,
Casula
,
O.
, and
Prada
,
C.
,
2013
, “
Data Set Reduction for Ultrasonic TFM Imaging Using the Effective Aperture Approach and Virtual Sources
,”
J. Phys. Conf. Ser.
,
457
(
1
), p.
012007
.
28.
Hunter
,
A. J.
,
Drinkwater
,
B. W.
, and
Wilcox
,
P. D.
,
2008
, “
The Wavenumber Algorithm for Full-Matrix Imaging Using an Ultrasonic Array
,”
IEEE Trans. Ultrason. Ferroelectr. Freq. Control
,
55
(
11
), pp.
2450
2462
.
29.
Zhang
,
J.
,
Drinkwater
,
B. W.
,
Wilcox
,
P. D.
, and
Hunter
,
A. J.
,
2010
, “
Defect Detection Using Ultrasonic Arrays: The Multi-Mode Total Focusing Method
,”
NDT&E Int.
,
43
(
2
), pp.
123
133
.
30.
Wilcox
,
P. D.
,
Holmes
,
C.
, and
Drinkwater
,
B. W.
,
2007
, “
Advanced Reflector Characterization With Ultrasonic Phased Arrays in NDE Applications
,”
IEEE Trans. Ultrason. Ferroelectr. Freq. Control
,
54
(
8
), pp.
1541
1550
.
31.
Holmes
,
C.
,
Drinkwater
,
B. W.
, and
Wilcox
,
P. D.
,
2008
, “
Advanced Post-Processing for Scanned Ultrasonic Arrays: Application to Defect Detection and Classification in Non-Destructive Evaluation
,”
Ultrasonics
,
48
(
6–7
), pp.
636
642
.
32.
Zhang
,
J.
,
Drinkwater
,
B. W.
, and
Wilcox
,
P. D.
,
2014
, “
Efficient Immersion Imaging of Components With Nonplanar Surfaces
,”
IEEE Trans. Ultrason. Ferroelectr. Freq. Control
,
61
(
8
), pp.
1284
1295
.
33.
Aschy
,
A.
,
Terrien
,
N.
,
Robert
,
S.
, and
Bentahar
,
M.
,
2017
, “
Enhancement of the Total Focusing Method Imaging for Immersion Testing of Anisotropic Carbon Fiber Composite Structures
,”
AIP Conf. Proc
,
1806
(
1
), p.
040005
.
34.
Mineo
,
C.
,
Lines
,
D.
, and
Cerniglia
,
D.
,
2021
, “
Generalised Bisection Method for Optimum Ultrasonic ray Tracing and Focusing in Multi-Layered Structures
,”
Ultrasonics
,
111
(
1
), p.
106330
.
35.
Zhang
,
J.
,
Drinkwater
,
B. W.
, and
Wilcox
,
P. D.
,
2013
, “
Comparison of Ultrasonic Array Imaging Algorithms for Nondestructive Evaluation
,”
IEEE Trans. Ultrason. Ferroelectr. Freq. Control
,
60
(
8
), pp.
1732
1745
.
36.
Hopkins
,
D.
,
Neau
,
G.
, and
Le Ber
,
L.
,
2011
, “
Advanced Phased-Array Technologies for Ultrasonic Inspection of Complex Composite Parts
,”
Proc., Smart Materials, Structures, & NDT in Aerospace
,
Montreal, CA
,
Nov. 2–4
.
37.
Davì
,
S.
,
Mineo
,
C.
,
MacLeod
,
C.
,
Pierce
,
S. G.
,
Gachagan
,
A.
,
Paton
,
S.
,
Munro
,
G.
,
O'Brien-O'Reilly
,
J.
, and
McCubbin
,
C.
,
2020
, “
Correction of B-Scan Distortion for Optimum Ultrasonic Imaging of Backwalls With Complex Geometries
,”
Insight-Non-Destr. Test. Cond. Monit.
,
62
(
4
), pp.
184
191
.
38.
Russell
,
J.
,
Long
,
R.
,
Duxbury
,
D.
, and
Cawley
,
P.
,
2012
, “
Development and Implementation of a Membrane-Coupled Conformable Array Transducer for Use in the Nuclear Industry
,”
Insight-Non-Destr. Test. Cond. Monit.
,
54
(
7
), pp.
386
393
.
39.
Nakahata
,
K.
,
Tokumasu
,
S.
,
Sakai
,
A.
,
Iwata
,
Y.
,
Ohira
,
K.
, and
Ogura
,
Y.
,
2016
, “
Ultrasonic Imaging Using Signal Post-Processing for a Flexible Array Transducer
,”
NDT&E Int.
,
82
(
1
), pp.
13
25
.
40.
Deutsch
,
W. A. K.
,
2000
, “
Automated Ultrasonic Inspection
,”
Proceedings of the WCNDT Conference
,
Rome, Italy
,
Oct. 15–21
.
41.
Zhang
,
J.
,
Cho
,
Y.
,
Kim
,
J.
,
Malikov
,
A. K. U.
,
Kim
,
Y. H.
,
Yi
,
J.-H.
, and
Li
,
W.
,
2021
, “
Non-Destructive Evaluation of Coating Thickness Using Water Immersion Ultrasonic Testing
,”
Coatings
,
11
(
11
), p.
1421
.
42.
Le Jeune
,
L.
,
Robert
,
S.
,
Dumas
,
P.
,
Membre
,
A.
, and
Prada
,
C.
,
2015
, “
Adaptive Ultrasonic Imaging With the Total Focusing Method for Inspection of Complex Components Immersed in Water
,”
AIP Conf. Proc
,
1650
(
1
), pp.
1037
1046
.
43.
Parrilla
,
M.
,
Brizuela
,
J.
,
Camacho
,
J.
,
Ibanez
,
A.
,
Nevado
,
P.
, and
Fritsch
,
C.
,
2008
, “
Dynamic Focusing Through Arbitrary Geometry Interfaces
,”
2008 IEEE Ultrasonics Symposium
,
Beijing, China
,
Nov. 2–5
.
44.
McKee
,
J. G.
,
Bevan
,
R. L.
,
Wilcox
,
P. D.
, and
Malkin
,
R. E.
,
2020
, “
Volumetric Imaging Through a Doubly-Curved Surface Using a 2D Phased Array
,”
NDT&E Int.
,
113
(
1
), p.
102260
.
45.
Kerr
,
W.
,
Pierce
,
S. G.
, and
Rowe
,
P.
,
2016
, “
Investigation of Synthetic Aperture Methods in Ultrasound Surface Imaging Using Elementary Surface Types
,”
Ultrasonics
,
72
(
1
), pp.
165
176
.
46.
Le Bihan
,
Y.
,
Joubert
,
P. Y.
, and
Placko
,
D.
,
2001
, “
Wall Thickness Evaluation of Single-Crystal Hollow Blades by Eddy Current Sensor
,”
NDT&E Int.
,
34
(
5
), pp.
363
368
.
47.
Hastie
,
S.
,
Chan
,
A.
,
Wiens
,
K.
,
Nagy
,
D.
,
Tollett
,
R.
, and
Lowden
,
P.
,
2021
, “
Computed Tomography Wall Thickness Inspection to Support Gas Turbine Blade Life Extension
,”
Proceedings of the ASME Turbo Expo 2021: Turbomachinery Technical Conference and Exposition. Volume 7: Industrial and Cogeneration; Manufacturing Materials and Metallurgy
,
7
(
1
), p.
V007T17A021
.
48.
Goldammer
,
M.
, and
Heinrich
,
W.
,
2006
, “
Active Thermography for Dimensional Measurements on gas Turbine Components
,”
9th European Conference on NDT
,
Berlin, Germany
,
Sept. 25–29
.
49.
Velichko
,
A.
, and
Wilcox
,
P. D.
,
2010
, “
An Analytical Comparison of Ultrasonic Array Imaging Algorithms
,”
J. Acoust. Soc. Am.
,
127
(
4
), pp.
2377
2384
.
50.
Hartigan
,
J. A.
, and
Wong
,
M. A.
,
1979
, “
Algorithm AS 136: A K-Means Clustering Algorithm
,”
J. R. Stat. Soc. C
,
28
(
1
), pp.
100
108
.