OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
Thermal Detection of Small Cracks in a Plate
Kurt Bryan1
Rose-Hulman Institute of Technology
May 20, 2011
1With some helpful discussions with Chris Earls, Cornell University School ofCivil and Environmental Engineering.
Kurt Bryan Thermal Detection of Small Cracks in a Plate
OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
1 Physical Setting
2 Forward Problem
3 Inverse Problem ApproachGaussian Test FunctionsEstimating the Crack ParametersEstimating Crack LengthNoise and Quantization ErrorNear Crack Expansion
4 Further Work
Kurt Bryan Thermal Detection of Small Cracks in a Plate
OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
Motivation
The USS Independence—newhull design, aluminum, proneto cracks.
How can we find small cracksefficiently?
Kurt Bryan Thermal Detection of Small Cracks in a Plate
OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
Motivation
The USS Independence—newhull design, aluminum, proneto cracks.
How can we find small cracksefficiently?
Kurt Bryan Thermal Detection of Small Cracks in a Plate
OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
Physical Setting
Experimental set-up:
Kurt Bryan Thermal Detection of Small Cracks in a Plate
OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
Physical Setting
Typical parameters: aluminum plate, 10 cm by 10 cm, 1 mmthickness, 5 to 20 watt (or more) laser.
Illuminate sample plate with laser for 0 < t < T (T from 30to 180 seconds)
Measure temperature on 250× 250 pixel grid, 0.1 degree Cresolution, every few seconds.
Goal: find any cracks—-position, orientation, size.
Kurt Bryan Thermal Detection of Small Cracks in a Plate
OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
Physical Setting
Typical parameters: aluminum plate, 10 cm by 10 cm, 1 mmthickness, 5 to 20 watt (or more) laser.
Illuminate sample plate with laser for 0 < t < T (T from 30to 180 seconds)
Measure temperature on 250× 250 pixel grid, 0.1 degree Cresolution, every few seconds.
Goal: find any cracks—-position, orientation, size.
Kurt Bryan Thermal Detection of Small Cracks in a Plate
OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
Physical Setting
Typical parameters: aluminum plate, 10 cm by 10 cm, 1 mmthickness, 5 to 20 watt (or more) laser.
Illuminate sample plate with laser for 0 < t < T (T from 30to 180 seconds)
Measure temperature on 250× 250 pixel grid, 0.1 degree Cresolution, every few seconds.
Goal: find any cracks—-position, orientation, size.
Kurt Bryan Thermal Detection of Small Cracks in a Plate
OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
Physical Setting
Typical parameters: aluminum plate, 10 cm by 10 cm, 1 mmthickness, 5 to 20 watt (or more) laser.
Illuminate sample plate with laser for 0 < t < T (T from 30to 180 seconds)
Measure temperature on 250× 250 pixel grid, 0.1 degree Cresolution, every few seconds.
Goal: find any cracks—-position, orientation, size.
Kurt Bryan Thermal Detection of Small Cracks in a Plate
OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
Heat Equation
Let x = (x1, x2) position, t time, Ω the (2D) “plate,” γ a crack inthe plate, u(x , t) denote plate temperature. Assume u satisfies
cρut − α4 u = f on Ω \ γ × (0,T )
α∂u
∂n= 0 on γ × (0,T )
α∂u
∂n= 0 on ∂Ω× (0,T )
u(x , 0) = 0.
Here f embodies the input heat source, possibly radiation orconvection losses through the plate face.
Kurt Bryan Thermal Detection of Small Cracks in a Plate
OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
Rescaled Heat Equation
After rescaling to Ω = (0, 1)2 and α/(cρ) = 1 we have
ut −4u = f on Ω \ γ × (0,T )
∂u
∂n= 0 on γ × (0,T )
∂u
∂n= 0 on ∂Ω× (0,T )
u(x , 0) = 0.
Kurt Bryan Thermal Detection of Small Cracks in a Plate
OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
A Typical Solution
The temperature u is discontinuous across the crack γ.
Kurt Bryan Thermal Detection of Small Cracks in a Plate
OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
What’s So Hard About Finding Cracks?
We want to find cracks that are small—on the order of a couplepixels long, or even less than one pixel!
There are 3 cracks in the field of view below:
Kurt Bryan Thermal Detection of Small Cracks in a Plate
OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
Gaussian Test FunctionsEstimating the Crack ParametersEstimating Crack LengthNoise and Quantization ErrorNear Crack Expansion
Integrate by Parts
Let φ(x , t) be a suitably smooth “test function”, and defineF = φt +4φ. A little integration by parts shows that∫ T
0
∫γ
[u]∂φ
∂nds dt =
∫Ω\γ
u(x ,T )φ(x ,T ) dx
−∫ T
0
∫Ω\γ
uF dx dt +
∫ T
0
∫∂Ω
u∂φ
∂nds dt
−∫ T
0
∫Ω\γ
f φ dx dt.
where [u] denotes the jump in u over γ in the direction n.
Everything on the right is computable from knowledge of u on(Ω \ γ)× (0,T ).
Kurt Bryan Thermal Detection of Small Cracks in a Plate
OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
Gaussian Test FunctionsEstimating the Crack ParametersEstimating Crack LengthNoise and Quantization ErrorNear Crack Expansion
Integrate by Parts
Let φ(x , t) be a suitably smooth “test function”, and defineF = φt +4φ. A little integration by parts shows that∫ T
0
∫γ
[u]∂φ
∂nds dt =
∫Ω\γ
u(x ,T )φ(x ,T ) dx
−∫ T
0
∫Ω\γ
uF dx dt +
∫ T
0
∫∂Ω
u∂φ
∂nds dt
−∫ T
0
∫Ω\γ
f φ dx dt.
where [u] denotes the jump in u over γ in the direction n.
Everything on the right is computable from knowledge of u on(Ω \ γ)× (0,T ).
Kurt Bryan Thermal Detection of Small Cracks in a Plate
OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
Gaussian Test FunctionsEstimating the Crack ParametersEstimating Crack LengthNoise and Quantization ErrorNear Crack Expansion
In Summary
We have the identity∫ T
0
∫γ
[u]∂φ
∂nds dt = q(φ)︸︷︷︸
computable from data
for any smooth function φ, where q(φ) is computable fromknowledge of u on (Ω \ γ)× (0,T ).
Can we use this, with suitable choices for φ, to deduce γ?
Kurt Bryan Thermal Detection of Small Cracks in a Plate
OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
Gaussian Test FunctionsEstimating the Crack ParametersEstimating Crack LengthNoise and Quantization ErrorNear Crack Expansion
In Summary
We have the identity∫ T
0
∫γ
[u]∂φ
∂nds dt = q(φ)︸︷︷︸
computable from data
for any smooth function φ, where q(φ) is computable fromknowledge of u on (Ω \ γ)× (0,T ).
Can we use this, with suitable choices for φ, to deduce γ?
Kurt Bryan Thermal Detection of Small Cracks in a Plate
OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
Gaussian Test FunctionsEstimating the Crack ParametersEstimating Crack LengthNoise and Quantization ErrorNear Crack Expansion
Assumptions
In what follows we’ll assume that
The crack γ is a line segment with center p = (p1, p2), angleθ, length |γ|.
The crack is “short.”
Kurt Bryan Thermal Detection of Small Cracks in a Plate
OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
Gaussian Test FunctionsEstimating the Crack ParametersEstimating Crack LengthNoise and Quantization ErrorNear Crack Expansion
Assumptions
In what follows we’ll assume that
The crack γ is a line segment with center p = (p1, p2), angleθ, length |γ|.The crack is “short.”
Kurt Bryan Thermal Detection of Small Cracks in a Plate
OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
Gaussian Test FunctionsEstimating the Crack ParametersEstimating Crack LengthNoise and Quantization ErrorNear Crack Expansion
Gaussian Test Functions
Consider a two-parameter family of test functions
φ(x1, x2) =e−((x1−a)2+(x2−b)2)/β2
πβ2
for some fixed β > 0. We can write∫ T
0
∫γ
[u]∂φ
∂nds dt = q(a, b)︸ ︷︷ ︸
computable from data
If γ is short we may approximate
q(a, b) ≈ ∂φ
∂n(p)
∫ T
0
∫γ
[u] ds dt.
Kurt Bryan Thermal Detection of Small Cracks in a Plate
OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
Gaussian Test FunctionsEstimating the Crack ParametersEstimating Crack LengthNoise and Quantization ErrorNear Crack Expansion
Observations
From
q(a, b) ≈ ∇φ(p) · n∫ T
0
∫γ
[u] ds dt.
we see thatq(a, b) ≡ 0
if no crack is present.
In fact, since φ effectively vanishes outside a ball of radius 2β,q(a, b) ≈ 0 if there’s no crack in a ball of radius 2β around (a, b).
Kurt Bryan Thermal Detection of Small Cracks in a Plate
OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
Gaussian Test FunctionsEstimating the Crack ParametersEstimating Crack LengthNoise and Quantization ErrorNear Crack Expansion
Observations
From
q(a, b) ≈ ∇φ(p) · n∫ T
0
∫γ
[u] ds dt.
we see thatq(a, b) ≡ 0
if no crack is present.
In fact, since φ effectively vanishes outside a ball of radius 2β,q(a, b) ≈ 0 if there’s no crack in a ball of radius 2β around (a, b).
Kurt Bryan Thermal Detection of Small Cracks in a Plate
OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
Gaussian Test FunctionsEstimating the Crack ParametersEstimating Crack LengthNoise and Quantization ErrorNear Crack Expansion
Raw Temperature Data and Function q
Time T = 3, raw temperature on the left, function q(a, b) on theright.
Kurt Bryan Thermal Detection of Small Cracks in a Plate
OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
Gaussian Test FunctionsEstimating the Crack ParametersEstimating Crack LengthNoise and Quantization ErrorNear Crack Expansion
Close Up View of Function q near Crack
Kurt Bryan Thermal Detection of Small Cracks in a Plate
OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
Gaussian Test FunctionsEstimating the Crack ParametersEstimating Crack LengthNoise and Quantization ErrorNear Crack Expansion
More Observations
From
q(a, b) ≈ ∇φ(p) · n∫ T
0
∫γ
[u] ds dt.
we see that
q ≡ 0 on the line containing the crack.
q attains its maximum/minimum orthogonally off the crack
center p, at points p±√
22 βn.
We can thus recover the crack center and angle (hence normal n)from q(a, b) for values of (a, b) near the crack.
Kurt Bryan Thermal Detection of Small Cracks in a Plate
OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
Gaussian Test FunctionsEstimating the Crack ParametersEstimating Crack LengthNoise and Quantization ErrorNear Crack Expansion
More Observations
From
q(a, b) ≈ ∇φ(p) · n∫ T
0
∫γ
[u] ds dt.
we see that
q ≡ 0 on the line containing the crack.
q attains its maximum/minimum orthogonally off the crack
center p, at points p±√
22 βn.
We can thus recover the crack center and angle (hence normal n)from q(a, b) for values of (a, b) near the crack.
Kurt Bryan Thermal Detection of Small Cracks in a Plate
OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
Gaussian Test FunctionsEstimating the Crack ParametersEstimating Crack LengthNoise and Quantization ErrorNear Crack Expansion
Crack Length
One can prove that∫ T
0
∫γ
[u] ds dt =π
4|γ|2
∫ T
0∇u0(p, t) · n dt + O(|γ|3).
where u0 is the solution to the heat equation on Ω (no crack).
Asa result
q(a, b) ≈ π
4|γ|2(∇φ(p) · n)
∫ T
0∇u0(p, t) · n dt.
Since we now know p and n, we can estimate |γ|.
Kurt Bryan Thermal Detection of Small Cracks in a Plate
OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
Gaussian Test FunctionsEstimating the Crack ParametersEstimating Crack LengthNoise and Quantization ErrorNear Crack Expansion
Crack Length
One can prove that∫ T
0
∫γ
[u] ds dt =π
4|γ|2
∫ T
0∇u0(p, t) · n dt + O(|γ|3).
where u0 is the solution to the heat equation on Ω (no crack). Asa result
q(a, b) ≈ π
4|γ|2(∇φ(p) · n)
∫ T
0∇u0(p, t) · n dt.
Since we now know p and n, we can estimate |γ|.
Kurt Bryan Thermal Detection of Small Cracks in a Plate
OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
Gaussian Test FunctionsEstimating the Crack ParametersEstimating Crack LengthNoise and Quantization ErrorNear Crack Expansion
Close Up View of Function q
True crack center is (0.7, 0.543), angle −60 degrees, length 0.01.Recovered parameters are (0.7, 0.544), angle −60.3 degrees, length0.0095 (no noise/quantization).
Kurt Bryan Thermal Detection of Small Cracks in a Plate
OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
Gaussian Test FunctionsEstimating the Crack ParametersEstimating Crack LengthNoise and Quantization ErrorNear Crack Expansion
Multiple Cracks
Power input 5 watts, real time 97.2 seconds, crack lengths0.02, 0.01, 0.005.
Kurt Bryan Thermal Detection of Small Cracks in a Plate
OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
Gaussian Test FunctionsEstimating the Crack ParametersEstimating Crack LengthNoise and Quantization ErrorNear Crack Expansion
Connection to Small Volume Expansions
The formula q(a, b) ≈ π4 |γ|
2(∇φ(p) · n)∫ T
0 ∇u0(p, t) · n dt can bewritten
q(a, b) ≈ π
4|γ|2
∫ T
0∇φ(p)tM∇u0(p, t) dt
with M = nnt and viewed as an extreme limiting case of thesmall-volume asymptotic expansions of Ammari, Kang, Vogelius,etc.
Physically interpretation: a crack looks like a “thermal dipole” atposition p, orientation n, dipole moment π
4 |γ|2∫ T
0 ∇u0 · n dt.
Kurt Bryan Thermal Detection of Small Cracks in a Plate
OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
Gaussian Test FunctionsEstimating the Crack ParametersEstimating Crack LengthNoise and Quantization ErrorNear Crack Expansion
Connection to Small Volume Expansions
The formula q(a, b) ≈ π4 |γ|
2(∇φ(p) · n)∫ T
0 ∇u0(p, t) · n dt can bewritten
q(a, b) ≈ π
4|γ|2
∫ T
0∇φ(p)tM∇u0(p, t) dt
with M = nnt and viewed as an extreme limiting case of thesmall-volume asymptotic expansions of Ammari, Kang, Vogelius,etc.
Physically interpretation: a crack looks like a “thermal dipole” atposition p, orientation n, dipole moment π
4 |γ|2∫ T
0 ∇u0 · n dt.
Kurt Bryan Thermal Detection of Small Cracks in a Plate
OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
Gaussian Test FunctionsEstimating the Crack ParametersEstimating Crack LengthNoise and Quantization ErrorNear Crack Expansion
Noise and Quantization Error
Infrared cameras used in this type of work have noise andquantization error, typically in the ballpark of 0.1 degrees C.
Power input 5 watts, real time 97.2 seconds, crack lengths0.02, 0.01, 0.005.
Kurt Bryan Thermal Detection of Small Cracks in a Plate
OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
Gaussian Test FunctionsEstimating the Crack ParametersEstimating Crack LengthNoise and Quantization ErrorNear Crack Expansion
Noise and Quantization Error
If we boost the power input to 25 watts (using more of thedynamic range of the camera) the results are
Power input 25 watts, real time 97.2 seconds, crack lengths0.02, 0.01, 0.005.
Kurt Bryan Thermal Detection of Small Cracks in a Plate
OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
Gaussian Test FunctionsEstimating the Crack ParametersEstimating Crack LengthNoise and Quantization ErrorNear Crack Expansion
Noise and Quantization Error
True crack parameters
center angle (degrees) length
(0.53, 0.42443) 0 0.005(0.75113, 0.4) 90 0.02(0.7, 0.54325) −60 0.01
Recovered parameters
center angle (degrees) length
(0.536, 0.429) −23.5 0.01(0.75, 0.4) 86.3 0.028(0.699, 0.543) −59.4 0.0168
Kurt Bryan Thermal Detection of Small Cracks in a Plate
OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
Gaussian Test FunctionsEstimating the Crack ParametersEstimating Crack LengthNoise and Quantization ErrorNear Crack Expansion
Near Field Expansion
The formula for q(a, b) was based on a “far field” approximation,∫ T
0
∫γ
[u]∂φ
∂nds dt ≈ ∇φ(p) · n
∫ T
0
∫γ
[u] ds dt
≈ ∇φ(p) · n(π
4|γ|2
∫ T
0∇u0(p, t) · n dt
)obtained by treating ∇φ as constant over the crack.
But we’re taking the Gaussian test function with center near thecrack. Is this approximation reasonable?
Kurt Bryan Thermal Detection of Small Cracks in a Plate
OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
Gaussian Test FunctionsEstimating the Crack ParametersEstimating Crack LengthNoise and Quantization ErrorNear Crack Expansion
Near Field Expansion
The formula for q(a, b) was based on a “far field” approximation,∫ T
0
∫γ
[u]∂φ
∂nds dt ≈ ∇φ(p) · n
∫ T
0
∫γ
[u] ds dt
≈ ∇φ(p) · n(π
4|γ|2
∫ T
0∇u0(p, t) · n dt
)obtained by treating ∇φ as constant over the crack.
But we’re taking the Gaussian test function with center near thecrack. Is this approximation reasonable?
Kurt Bryan Thermal Detection of Small Cracks in a Plate
OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
Gaussian Test FunctionsEstimating the Crack ParametersEstimating Crack LengthNoise and Quantization ErrorNear Crack Expansion
Near Field Expansion
Kurt Bryan Thermal Detection of Small Cracks in a Plate
OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
Gaussian Test FunctionsEstimating the Crack ParametersEstimating Crack LengthNoise and Quantization ErrorNear Crack Expansion
Advantages
Computationally very fast.
Robust against model departures—precise knowledge of powerinput/profile, emissivity, boundary conditions not necessary.Even convection/radiation makes little difference.
Flexibility—test functions can be tailored to incorporate mostrelevant data.
Physical insight—illuminates what variables/geometries aremore important.
Kurt Bryan Thermal Detection of Small Cracks in a Plate
OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
Gaussian Test FunctionsEstimating the Crack ParametersEstimating Crack LengthNoise and Quantization ErrorNear Crack Expansion
Advantages
Computationally very fast.
Robust against model departures—precise knowledge of powerinput/profile, emissivity, boundary conditions not necessary.Even convection/radiation makes little difference.
Flexibility—test functions can be tailored to incorporate mostrelevant data.
Physical insight—illuminates what variables/geometries aremore important.
Kurt Bryan Thermal Detection of Small Cracks in a Plate
OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
Gaussian Test FunctionsEstimating the Crack ParametersEstimating Crack LengthNoise and Quantization ErrorNear Crack Expansion
Advantages
Computationally very fast.
Robust against model departures—precise knowledge of powerinput/profile, emissivity, boundary conditions not necessary.Even convection/radiation makes little difference.
Flexibility—test functions can be tailored to incorporate mostrelevant data.
Physical insight—illuminates what variables/geometries aremore important.
Kurt Bryan Thermal Detection of Small Cracks in a Plate
OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
Gaussian Test FunctionsEstimating the Crack ParametersEstimating Crack LengthNoise and Quantization ErrorNear Crack Expansion
Advantages
Computationally very fast.
Robust against model departures—precise knowledge of powerinput/profile, emissivity, boundary conditions not necessary.Even convection/radiation makes little difference.
Flexibility—test functions can be tailored to incorporate mostrelevant data.
Physical insight—illuminates what variables/geometries aremore important.
Kurt Bryan Thermal Detection of Small Cracks in a Plate
OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
Further Work
Work out the near crack expansion—better length estimates?
For cracks near or below one pixel, would a careful model ofthe image capture process (point-spread function, sensorgeometry) be worth it?
A more realistic boundary condition on the crack is∂u/∂n = k[u]. In the elliptic case the asymptotics are∫
γ[u] ds =
π/4
1 + 83πk |γ|
|γ|2 + O(c(k)|γ|3).
This ought to hold in the parabolic case too.
Would other test functions yield better results?
Kurt Bryan Thermal Detection of Small Cracks in a Plate
OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
Further Work
Work out the near crack expansion—better length estimates?
For cracks near or below one pixel, would a careful model ofthe image capture process (point-spread function, sensorgeometry) be worth it?
A more realistic boundary condition on the crack is∂u/∂n = k[u]. In the elliptic case the asymptotics are∫
γ[u] ds =
π/4
1 + 83πk |γ|
|γ|2 + O(c(k)|γ|3).
This ought to hold in the parabolic case too.
Would other test functions yield better results?
Kurt Bryan Thermal Detection of Small Cracks in a Plate
OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
Further Work
Work out the near crack expansion—better length estimates?
For cracks near or below one pixel, would a careful model ofthe image capture process (point-spread function, sensorgeometry) be worth it?
A more realistic boundary condition on the crack is∂u/∂n = k[u]. In the elliptic case the asymptotics are∫
γ[u] ds =
π/4
1 + 83πk |γ|
|γ|2 + O(c(k)|γ|3).
This ought to hold in the parabolic case too.
Would other test functions yield better results?
Kurt Bryan Thermal Detection of Small Cracks in a Plate
OutlinePhysical Setting
Forward ProblemInverse Problem Approach
Further Work
Further Work
Work out the near crack expansion—better length estimates?
For cracks near or below one pixel, would a careful model ofthe image capture process (point-spread function, sensorgeometry) be worth it?
A more realistic boundary condition on the crack is∂u/∂n = k[u]. In the elliptic case the asymptotics are∫
γ[u] ds =
π/4
1 + 83πk |γ|
|γ|2 + O(c(k)|γ|3).
This ought to hold in the parabolic case too.
Would other test functions yield better results?
Kurt Bryan Thermal Detection of Small Cracks in a Plate