US 12,241,870 B2
Ultrasonic non-destructive test method and system using deep learning, and auto-encoder-based prediction model training method used therefor
Jong Moon Ha, Daejeon (KR); Won Jae Choi, Daejeon (KR); and Hong Min Seung, Daejeon (KR)
Assigned to KOREA RESEARCH INSTITUTE OF STANDARDS AND SCIENCE, Daejeon (KR)
Appl. No. 17/776,909
Filed by KOREA RESEARCH INSTITUTE OF STANDARDS AND SCIENCE, Daejeon (KR)
PCT Filed Apr. 4, 2022, PCT No. PCT/KR2022/004811
§ 371(c)(1), (2) Date May 13, 2022,
PCT Pub. No. WO2022/234957, PCT Pub. Date Nov. 10, 2022.
Claims priority of application No. 10-2021-0057220 (KR), filed on May 3, 2021.
Prior Publication US 2024/0053302 A1, Feb. 15, 2024
Int. Cl. G01N 29/44 (2006.01); G01N 29/024 (2006.01); G06N 3/0455 (2023.01); G06N 3/08 (2023.01)
CPC G01N 29/4472 (2013.01) [G01N 29/024 (2013.01); G06N 3/0455 (2023.01); G06N 3/08 (2013.01); G01N 2291/011 (2013.01); G01N 2291/023 (2013.01); G01N 2291/0289 (2013.01); G01N 2291/101 (2013.01)] 8 Claims
OG exemplary drawing
 
1. A method for training an autoencoder-based prediction model used in an ultrasonic NDT (Non-Destructive Test) method using deep learning, the method comprising:
an ultrasonic signal acquisition step of acquiring a normal signal by transmitting an ultrasonic wave to a test object with no defect, and receiving an ultrasonic wave reflected from the test object;
a prediction model training step of training a prediction model through a process of minimizing a loss function based on Equation 1 below by using the normal signal:
L(xn)=∥xn−gψ(fϕ(xn))∥2  Equation 1,
where xn represents a measured signal, ψ and ϕ represent training parameters, fϕ represents a transfer function of an encoder, and gψ represents a transfer function of a decoder;
an ultrasonic signal reacquisition step of acquiring a remeasured signal including a pseudo-normal signal for a portion with no defect and a defect signal for a portion with a defect by transmitting/receiving an ultrasonic wave to/from a test object with a defect;
a pseudo-normal signal extraction step of extracting only the pseudo-normal signal from the remeasured signal; and
a prediction model retraining step of retraining the prediction model through a process of minimizing a loss function based on Equation 2 below by using the normal signal and the pseudo-normal signal:
L(xncustom character)=∥xn−gψre(fϕre(xn))∥2+∥custom character−gψre(fϕre(custom character))∥2  Equation 2,
where xn represents the measured signal, custom character represents the remeasured signal, ψre and ϕre represent retraining parameters, fϕre represents a transfer function of the encoder, and gψre represents a transfer function of the decoder;
wherein the pseudo-normal signal extraction step further comprises:
a MAD (Mean Absolute Difference) calculation step of calculating an MAD by averaging absolute values of differences between the normal signal and remeasured signals;
a threshold calculation step of calculating a threshold based on Equation 3 below by using a distribution of the MADs; and
a pseudo-normal signal determination step of determining that a remeasured signal is the pseudo-normal signal, when the MAD is smaller than the threshold,
wherein the MAD indicates how much the remeasured signals differ from the normal signal:
threshold=μMAD(1)+ασMAD(1)  Equation 3,
where μMAD(1) and ασMAD(1) represent an average and standard deviation of a first Gauss distribution of MADs estimated by a Gaussian mixture model, and α represents a critical parameter.