Method and system for training an artificial neural network utilizing physics based knowledge
View Patent ↗A method for training an artificial neural network including classifiers for material characterization. The method includes obtaining functional data having phase and amplitude, registering functional data by phase-amplitude separation and statistical analysis on the phase-amplitude separated data with an elastic distance to produce aligned functional data, performing dimensional reduction on the aligned functional data to produce a dimensional representation of the functional space, performing, by a computer system, a training operation to train an artificial neural network based on the dimensional representation of the functional space. A method and system for material characterization is also disclosed.
1. A method for training an artificial neural network comprising:
obtaining functional data having phase and amplitude;
registering the functional data by phase-amplitude separation of the functional data to produce separated phase and amplitude components with an elastic distance and performing statistical analysis on the separated phase and amplitude components to produce aligned functional data;
performing dimensional reduction on the aligned functional data to produce a dimensional representation of a functional space of the aligned functional data; and
performing, by a computer system, a training operation to train an artificial neural network based on the dimensional representation of the functional space.
2. The method of claim 1 , wherein the registering functional data includes elastic function data alignment.
3. The method of claim 2 , wherein the elastic function data alignment includes applying a square root slope function.
4. The method of claim 1 , wherein the elastic distance is a mean function utilized to warp functions of the phase and amplitude of the functional data.
5. The method of claim 4 , wherein the elastic distance is defined as the following equation:
d
a
(
f
1
,
f
2
)
=
inf
γ
∈
Γ
q
1
-
(
q
2
◦γ
)
γ
.
where d a is the elastic distance, ƒ 1 , ƒ 2 are any two functions of the functional data, group Γ is a set of orientation-preserving diffeomorphism representing the phase of the functional data, q 1 and q 2 are the square root slope function of the functional data, γ is the amount of time-warping of ƒ 1 and ƒ 2 .
6. The method of claim 1 , wherein dimensional reduction on the aligned functional data is performed with uniform manifold approximation and projection.
7. The method of claim 1 , wherein the artificial neural network is a classifier for material characterization.
8. The method of claim 7 , wherein the classifier is a support vector machine (SVM), a random forest (RF) classifier or convolutional neural network (CNN).
9. A method for material characterization comprising:
training a classifier according to the method of claim 1 ;
scanning a sample with an x-ray computed tomography scanner;
obtaining functional data from the computed tomography scanner;
characterizing the material of the sample with the functional data from the computed tomography scanner; and
displaying the characterization result on a display.
10. A system for training a classifier for a material characterization comprising:
at least one processor; a non-transitory, computer-readable medium having instructions stored thereon that are executable by the at least one processor to cause the system to:
obtain the functional data having phase and amplitude;
register functional data by phase-amplitude separation of the functional data to produce separated phase and amplitude components with an elastic distance and perform statistical analysis on the separated phase and amplitude components to produce aligned functional data;
perform dimensional reduction on the aligned functional data to produce a dimensional representation of a functional space of the aligned functional data; and
train a classifier with the dimensional representation of the functional space.
11. The system of claim 10 , wherein the functional data is voxel data from an x-ray computed tomography scan.
12. The system of claim 10 , wherein the registering functional data includes elastic function data alignment.
13. The system of claim 10 , wherein the elastic distance is the mean function utilized to warp functions of the phase and amplitude.
14. The system of claim 13 , wherein the elastic distance is defined as the following equation:
d
a
(
f
1
,
f
2
)
=
inf
γ
∈
Γ
q
1
-
(
q
2
◦γ
)
γ
.
where d a is the elastic distance, ƒ 1 , ƒ 2 are any two functions of the functional data, group Γ is a set of orientation-preserving diffeomorphism representing the phase of the functional data, q 1 and q 2 are the square root slope function of the functional data, γ is the amount of time-warping of ƒ 1 and ƒ 2 .
15. The system of claim 10 , wherein dimensional reduction on the aligned functional data is performed with uniform manifold approximation and projection.
16. The system of claim 10 , wherein the classifier is one of a support vector machine (SVM), a random forest (RF) classifier or convolutional neural network (CNN).
17. The system of claim 10 , wherein the dimensional representation of the functional space is a fuzzy representation of the data manifold.
18. A system for material characterization comprising:
a computed tomography scanner;
a processor including classifier trained by the system of claim 10 ; and
a display for displaying the output of the processor to identify materials scanned by the computed tomography scanner.
19. The system of claim 18 , wherein the computed tomography scanner is a hyperspectral computed tomography scanner.
20. The system of claim 18 , wherein the system identifies materials with an accuracy greater than 30% more accurate than a process having a classifier trained utilizing functional data that is not registered and aligned with the elastic distance.