서지주요정보
신경망 학습 기반 레이저 및 위상배열 초음파 검사의 신호 분류 및 결함 평가 연구 = Signal Classification and Defect Evaluation in Laser and Phased-Array Ultrasonic Inspection based-on Training Neural Network
서명 / 저자 신경망 학습 기반 레이저 및 위상배열 초음파 검사의 신호 분류 및 결함 평가 연구 = Signal Classification and Defect Evaluation in Laser and Phased-Array Ultrasonic Inspection based-on Training Neural Network / 김용호.
발행사항 [대전 : 한국과학기술원, 2024].
Online Access 원문보기 원문인쇄

소장정보

등록번호

8042699

소장위치/청구기호

학술문화관(도서관)2층 학위논문

DAE 24002

휴대폰 전송

도서상태

이용가능(대출불가)

사유안내

반납예정일

리뷰정보

초록정보

In this study, through-transmission laser ultrasonic test and phased-array ultrasonic test, which are classified as advanced non-destructive testing methods, were selected to develop signal classification and defect evaluation technology based on training neural network. First, through-transmission laser ultrasonic test has great advantages for long-distance non-contact measurement, but it is more affected by noise than contact measurement, and often suffers from poor signal quality. Phased-array ultrasonic test has advantages over conventional ultrasonic test in terms of inspection range and accuracy by generating ultrasonic waves at various angles, but it has the problem that the process of accurately classifying suspicious normal signals and actual defect signals is complex and time-consuming. To solve these problems, this study proposes a method to minimize the influence of noise on the learning of ultrasonic signals by designing an appropriate neural network structure, and a method to train a neural network with domain knowledge such as welding shape along with ultrasonic signals. The proposed method actually performed well in removing noise from laser ultrasonic signals and contributed to improving the signal classification accuracy of the neural network by allowing features to be extracted from the actual signal rather than noise during training. In addition, four feature values were selected as domain knowledge for phased-array ultrasonic signals and trained together with the ultrasonic signals, and the signal classification accuracy and defect evaluation accuracy were greatly improved. First, an auto-encoder neural network was introduced as a method for removing noise, and it was trained to restore the de-noised signal instead of restoring the input signal as it was. In laser ultrasonic inspection, a large amount of ultrasonic signals are acquired due to a tight inspection interval, and this was used as an advantage to obtain a signal that minimizes the influence of noise by calculating the average signal with the surrounding signals. Then, the neural network was trained by setting the original signal as the input value and the noise-reduced signal as the target value. In addition, a new signal classification neural network was designed by recycling only the encoder part that extracts features from the trained auto-encoder network. Since the recycled encoder part was trained to extract features only form the actual ultrasonic signal, not noise, its advantages can be maximized when reused for signal classification problems. The proposed method was verified on four different CFRP structures, and images visualizing delamination, a typical defect of composite structures, were presented as the inspection results. Next, when it is difficult to accurately classify defects by learning only ultrasonic signals, a new training dataset and neural network structure were proposed as a way to reflect the inspector’s domain knowledge in training neural network. For weld inspection, two values representing the relative position between the defect and weld boundary, one value reflecting the characteristics of the ultrasonic signal itself, and one value considering the inspection environment were selected, and the dataset and neural network structure were redesigned so that the four selected value can be input and processed in parallel with the ultrasonic signal. As a result of learning the domain knowledge together, signal classification accuracy improved by more than 8% compared to learning only ultrasonic signals. This means that the signal classification accuracy, which had been around 90%, has improved to a practical level. The defect evaluation results of stainless steel and carbon steel welded specimens were compared with the expert evaluation results, and it was confirmed that the accuracy was not inferior to the expert level. Actually, the scope of this study was limited to two advanced ultrasonic inspection methods, but it is clear that the presented methodology and direction of this research can be fully expanded to other ultrasound-based inspections. In cases where the quality of the measured ultrasonic signals is poor, or the ultrasonic signals are insufficiently informative to train a neural network, or both, the direction of this research can be an excellent guide. Furthermore, it is expected that through continued study and expansion in the future, automated expert-level defect evaluation technology for general ultrasonic test will be developed.

본 연구에서는 초음파검사 방법 중 고등 비파괴검사 방법으로 분류되는 투과식 레이저 초음파 검사와 위상배열 초음파 검사를 선정하여 신경망 학습 기반의 신호 분류 및 결함 평가 기술을 개발하였다. 먼저, 투과식 레이저 초음파 검사는 원거리 비접촉 방식의 계측에 큰 장점이 있지만 접촉식 계측 방식보다는 잡음의 영향을 많이 받을 수밖에 없기 때문에 종종 신호 품질이 저하되는 문제가 있다. 위상배열 초음파 검사는 여러 각도로 초음파를 발생시켜 일반 초음파 검사보다 검사 범위 및 정확도에 장점이 있지만, 의심스러운 정상 신호와 실제 결함 신호를 정확히 분류하는 과정이 복잡하고 어렵다는 문제가 있다. 이러한 문제들을 해결하기 위해 본 연구에서는 적절한 신경망 구조를 설계하여 초음파 신호의 학습 시에 잡음의 영향을 최소화하는 방법과 용접 형상 등의 도메인 지식을 초음파 신호와 함께 신경망에 학습시키는 방법 등을 제안하였다. 제안된 방법은 실제로 레이저 초음파 신호에서 잡음을 제거하는 성능이 뛰어났고, 학습 시에 잡음이 아닌 실제 신호에서 특징이 추출될 수 있게 유도하여 신경망의 신호 분류 정확도 향상에 기여하였다. 그리고 위상배열 초음파 신호에서는 네 가지 물리량을 도메인 지식으로 선정하여 초음파 신호와 함께 학습시켰고, 신호 분류 정확도와 결함 평가 정확도가 크게 향상되었다. 먼저, 잡음을 제거하기 위한 방법으로 오토 인코더 신경망을 도입하였는데, 입력된 신호를 그대로 복원하는 것이 아니라 잡음이 제거된 신호를 복원할 수 있도록 학습을 유도하였다. 레이저 초음파 검사에서는 촘촘한 검사 간격으로 인해 방대한 양의 초음파 신호가 획득되는데, 이를 장점으로 활용하여 주변 신호와의 평균 계산을 통해 잡음의 영향을 최소화한 신호를 얻었다. 그리고 원 신호를 입력값으로, 잡음이 감소된 신호를 목표값으로 설정하여 신경망을 학습시켰고, 학습된 신경망에서 특징을 추출하는 부분인 인코더 부분을 재활용하여 초음파 신호 분류 신경망을 설계하였다. 재활용된 인코더 부분은 잡음이 아닌 실제 초음파 신호에서만 특징을 추출할 수 있도록 학습된 신경망이기 때문에 신호 분류 문제에 재활용할 시에 그 장점이 극대화될 수 있다. 네 가지 CFRP 구조물에서 위의 방법을 검증하였고, 복합재의 대표적인 결함인 층간 분리 등의 결함을 분류하여 가시화한 영상을 결과로 제시하였다. 다음으로, 초음파 신호의 학습만으로는 결함 신호를 정확히 분류하기 어려울 때, 검사자의 도메인 지식을 신경망 학습에 반영하기 위한 방법으로 새로운 학습 데이터셋과 신경망 구조를 제안하였다. 용접 검사에서는 결함 신호와 용접 경계면 사이의 상대적인 위치 관계를 나타내는 두 가지 값과 초음파 신호 자체적인 특성을 반영한 한 가지 값, 그리고 마지막으로 검사 환경을 고려한 한 가지 값을 선정하였고, 이러한 도메인 지식들이 초음파 신호와 병렬적으로 입력되어 처리될 수 있도록 데이터셋과 신경망 구조를 새롭게 설계하였다. 도메인 지식을 함께 학습시킨 결과 초음파 신호만 학습시킨 것에 비해 약 8% 이상의 신호 분류 정확도가 향상되었는데, 이는 약 90% 수준에서 개선되지 않던 분류 정확도를 실용적인 수준까지 끌어올린 것이었다. 그리고 스테인리스강과 탄소강 용접 시편에서 결함 평가 결과를 전문가 평가 결과와 비교하여 그 정확도가 전문가 수준에 비해 떨어지지 않는 것을 확인하였다. 본 연구에서는 두 가지 고등 초음파 검사 방법으로 연구 범위를 제한하였지만, 제시된 방법론이나 방향성은 초음파 기반의 타 검사 기법으로도 충분히 확장될 수 있음이 자명하다. 계측된 초음파 신호의 품질이 떨어지는 경우나 초음파 신호 자체만으로는 정보가 부족한 경우, 혹은 두 가지 문제가 모두 발생한 경우에 본 연구의 방향이 훌륭한 길잡이가 될 수 있다. 또한 향후 지속적인 연구 및 외연 확장을 통해 초음파 검사에 대한 전문가 수준의 결함 평가 자동화 기술이 개발될 수 있을 것으로 예상된다.

서지기타정보

서지기타정보
청구기호 {DAE 24002
형태사항 v, 85 p. : 삽도 ; 30 cm
언어 한국어
일반주기 저자명의 영문표기 : Yong-Ho Kim
지도교수의 한글표기 : 이정률
지도교수의 영문표기 : Jung-Ryul Lee
부록 수록
학위논문 학위논문(박사) - 한국과학기술원 : 항공우주공학과,
서지주기 참고문헌 : p. 83-85
주제 신경망 학습
레이저 초음파검사
위상배열 초음파검사
데이터 평가
진단 인공지능
Neural network training
Laser ultrasonic test
Phased-array ultrasonic test
Data evaluation
Diagnostic artificial intelligence
QR CODE

책소개

전체보기

목차

전체보기

이 주제의 인기대출도서

Comparison between UT and RT

Normal indications and defects on (a) C-scan and (b) B-scan images

S-scan images of (a) LF defect and (b) normal indication with overlaid welding boundary

S-scan images of (a) IP defect and (b) geometry signals with overlaid welding boundary

Surface and bulk waves generated by pulsed laser beam

Schematic configuration of LDV

Four types of inspection targets: (a) 2 mm thick CFRP plate, (b) 25 mm thick CFRP plate, (c) 4mm thick plastic and 22 mm thick CFRP curvature, and (d) 4 mm thick plastic, 22 mm thick CFRP, and 2 mm thick GFRP curvature

(a) Detailed structure and (b) top view of the 2 mm thick CFRP plate

Both sides of CFRP plate: (a) peel-ply side and (b) mold side

60 m long wind turbine blade spar cab structure

Information on artificial defects embedded in CFRP structure to simulate the delamination

Schematic configuration of TT UPI system for CFRP plate inspection

Cylindrical structure of 4 mm thick plastic and 22 mm thick CFRP

Schematic diagram of R TT UPI system for inspection of cylindrical structure

Defects map of the 26 mm thick cylindrical structure

Cylindrical structure of 4 mm thick plastic, 22 mm thick CFRP and 2 mm thick GFRP

Defects map of 28 mm thick cylindrical structure

How to simulate crack defects during carbon fiber filament winding process

Laser ultrasonic signals measured from (a) normal, (b) delamination, and (c) impact damage of the 2 mm thick CFRP plate

Comparison of 32-average laser ultrasonic signals from 2 mm thick CFRP plate

Randomly selected normal ultrasonic signals from 25 mm thick CFRP plate

Randomly selected defect ultrasonic signals from 25 mm thick CFRP plate

Randomly selected noise signals from support aluminum alloy

Comparison of 32-average laser ultrasonic signals from 25 mm thick CFRP plate

Ultrasonic wave propagationimage and three types of signals from 4 mm thick plastic and 22 mm thick CFRP structure

Randomly selected normal ultrasonic signals from 4 mm thick plastic and 22 mm thick CFRP structure

Randomly selected defect ultrasonic signals from 4 mm thick plastic and 22 mm thick CFRP structure

Randomly selected fiber debonding ultrasonic signals from 4 mm thick plastic and 22 mm thick CFRP structure

Comparison of 32-average laser ultrasonic signals from 4 mm thick plastic and 22 mm thick CFRP structure

Schematic workflow of data classification

Simple schematic of auto-encoder

Diagram of deep convolutional autoencoder

Layer structure of designed DCAE model

Proposal of DCAE training method using averaged signal as target signal

Comparison of input(raw) and target(averaged) signals (a) in normal area and (b) in defect area

Comparison ofinput and reconstructed signals (a) at normal point, (b) at defect point and (c) at support point

How to construct CNN structure using pre-trained deep convolutional encoder

Layer structure of CNN model based on pre-trained deep convolutional encoder

Size and configuration of datasets for CNN model training: (a) 2mm CFRP plate, (b) 25 mm CFRP plate, and (c) 4 mm plastic and 22 mm CFRP with/without 2 mm GFRP cylinder

Classification results for test datasets: (a) 2mm CFRP plate, (b) 25 mm CFRP plate, and (c) 4 mm plastic and 22 mm CFRP with/without 2 mm GFRP cylinder

Comparison of (a) UWPI freeze frame and (b) signal classification image on 2 mm thick CFRP plate

Comparison of (a) UWPI freeze frame and (b) signal classification image on 25 mm thick CFRP plate

Comparison of (a) UWPI freeze frame and (b) signal classification image on 4 mm thick plastic and 22 mm thick CFRP curvature structure

CT results for naturally occurring defect

CT results with improved contrast ratio

Comparison of (a) UWPI freeze frame and (b) signal classification image on 4 mm thick plastic, 22 mm thick CFRP, and 2 mm thick GFRP curvature structure

The overall process of PAUT data evaluation AI

Information about weld specimens

Inspection schematic for weld specimens

Data structure of the OPD extension file

Beam angle calculation through coordinate system conversion

A-scan plot results of (a) OmniPC software and

Description of the sectorial scan image

How to restore S-scan image from 2D array of A-scan signals

S-scan image results of (a) OmniPC software and

Parallel structure of neural network for simultaneous processing of

Training dataset structure

Cross validation results of how classification accuracy varies with the number of hidden layers

Comparison of classification accuracy between two neural networks

Confusion matrices for the same test dataset after training with

Defect evaluation results for PL27619 specimen

(a) Radiographic test result for PL27619 specimen and

Confusion matrices for defect evaluation results of