Preparation of Steel Samples with Corrosion Damage
A total of nine steel samples, each measuring 12 in. × 4 in.
(30.5 cm × 10.2 cm) were prepared for corrosion inspection
using the robotic ECA setup. The samples included three
distinct geometries: one flat, four concave, and four convex, as
shown in Figure 3. These shapes were achieved using a sheet
roller, with curvatures designed according to corrosion stan-
dards detailed in Table 1. After shaping, the samples underwent
a seven-day salt spray corrosion process in accordance with
ASTM B117-03 (ASTM 2003). To create a calibration area for the
ECA, the bottom inch (2.54 cm) of each sample was treated
with a corrosion inhibitor however, some corrosion leached
into this section. Residual rust was removed using a soaking
solution.
Comparison of Processed ECA Data with Microscope
Depth Profiling
Ground truth volumetric data of surface corrosion were
obtained using a high-resolution digital microscope for
comparison with the ECA data. To locate the defects,
2D cross-correlation was applied, in which the microscope
data was searched across the ECA data. The best matching
pattern was indicated by the maximum correlation value
between the two sets. The patterns from ECA voltages cor-
related well with the depths of the microscope data.
Challenges in Robotic ECA Inspection of Curved
Samples
Detecting fine surface or underpaint corrosion defects (e.g.,
50 µm deep) on curved metallic surfaces using robotic arms
is challenging due to strict liftoff, tilt, and thermal stability
requirements for the ECA probe. For example, ECT typically
necessitates maintaining a close and constant liftoff (e.g.,
1 mm) between the coil probe and the metallic surface. Small
liftoff ensures a noncontact testing method, maximizing the
induction of eddy currents and achieving a high signal-to-
noise ratio (SNR) for the measurements. Additionally, surface
corrosion detection often requires sensor coils to operate at
excitation frequencies in the MHz range. As a result, probe
misalignment and temperature fluctuations can introduce
errors during the scanning process, affecting both defect
detection and the accurate quantification of defect depth.
This section explores the challenges faced during robotic ECT
and presents typical experimental images obtained using the
proposed robotic NDE system with computer vision.
Effect of Inconsistent Probe Liftoff and Tilt on ECA
Measurements
Inconsistent liftoff typically introduces low spatial frequency
trends that overlap with defect indications, complicating image
processing and defect detection. Robotic manipulators can also
introduce inconsistent sensor tilt, affecting the distribution of
the excitation magnetic field from the coil sensor. For the rigid
and linear array probe, orientation errors affect each coil as a
body transformation, with coils at the edges of the array having
a different liftoff compared to the coils in its center. Improper
liftoff and/or tilt of the ECA probe can not only disrupt mea-
surements but also cause collisions with the test part, poten-
tially leading to immediate damage to the probe or long-term
degradation if dragged over rough surfaces.
B1 B2 B3 B4 C1 C2 C3 C4 A
Figure 3. Steel samples with corrosion damage: (a) flat sample A (b) concave samples B1–B4 (c) convex samples C1–C4.
TA B L E 1
Curvatures of test samples with corrosion damage
Sample A B1 B2 B3 B4 C1 C2 C3 C4
Curvature k (1/m) 0.000 0.304 0.397 0.527 0.842 0.304 0.451 0.528 0.842
A P R I L 2 0 2 5 • M AT E R I A L S E V A L U AT I O N 65

Liftoff and tilt inconsistencies often result from issues in
the reconstruction and registration of the virtual geometry for
path planning, such as misalignments between the virtual and
physical surface geometries. Additional factors include limita-
tions in the robot’s hardware and the synchronization between
the ECA data acquisition instrument and the robot. While
post-processing methods can help mitigate these errors and
improve defect detection, some defects may still be missed if
the errors exceed the algorithm’s corrective capacity.
Part Surface Reconstruction Error
Errors in the reconstructed surface of the test specimen can
arise from several sources. One common issue is the alignment
of the model, where misalignments lead to translation errors
along the scan path. If the model is not properly oriented with
its physical counterpart, rotation and liftoff variations will
occur throughout the entire scanning process. Another chal-
lenge is the discrepancy between the virtual model and the
physical surface geometry.
In this work, surface reconstruction of steel samples is
achieved using structured light with a CR-01 scanner offering
0.1 mm accuracy. However, metallic surfaces can reflect light,
causing inaccuracies during scanning, a problem that can be
mitigated using blue-light reconstruction (Zhan et al. 2015).
Additionally, errors may arise if the sample lies outside the
sensor’s effective range, affecting both translation and rotation.
Ray-plane intersection algorithms calculate translation and
rotation based on the mesh faces’ normal directions (Hamilton
et al. 2024).
Rotational errors between scan points can introduce dis-
crepancies, which can be mitigated using detrending algo-
rithms applied to the raster path. These errors arise at various
stages, as illustrated in Figures 4a and 4b from a scan of flat
Sample A (Figure 3a). Figure 4a depicts the delta rotation
acquired from the robot alone, revealing a baseline pathing
error. In raster pathing, certain scan lines shift more than
others due to orientation errors on the mesh. Figure 4b high-
lights spatial variations in the ECA signal caused by positioning
errors from pathing inaccuracies.
Mesh registration also introduces potential body transfor-
mation errors. The process, which relies on point-pair picking
(PPP), aligns the mesh within the robot’s reference frame.
Misalignment in laser calibration or displacement of calibra-
tion marks during sample placement can introduce errors,
as can human inaccuracies in selecting calibration points. To
reduce these errors, the root-mean-square (RMS) error was
minimized to approximately 1 mm. Although PPP provides
an initial approximation, body errors were further corrected
using 2D detrending. Figure 4c illustrates this correction with
four calibration markers (R0–R3), whose reconstructed posi-
tions were compared to physical locations identified via laser
calibration. The overall RMS error across all four points was
0.516 mm.
Robot Mastering Error
This error occurs during the calibration of a robotic arm’s
joints, where each joint requires precise alignment at zero
degrees. While state-of-the-art robots automate this process,
the robotic arm used in the presented system relies on manual
measurements and markings on the robot itself. Consequently,
sub-degree misalignments can occur at each joint. Although
these errors may seem minor, they result in more complex
orientation discrepancies within the robot’s joint system,
1010 1030
x (mm)
1050 1020 1040
—100
—50
R3 R0
R2 R1
0
50
100
—100
—50
0
50
100
0.9
1
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
x (mm)
0.3
0.2
0.1
0
—0.1
—0.2
Figure 4. Effects from mesh pathing errors: (a) changes of rotation
xyz of the probe during the scan of flat Sample A as provided from
the robot controller (b) corresponding voltage trend registered by
the ECA (c) aligned scanning environment from structured-light
reconstruction after selecting point pairs on the flat sample.
ME
|
ROBOTICECA
66
M AT E R I A L S E V A L U AT I O N • A P R I L 2 0 2 5
y
(mm)
y
(mm) Degree
V
(V)