지능형 용접 결함 분류를 위한 초음파 특징에 관한 연구

다국어 초록 (Multilingual Abstract)

Flaw classification is one of the fundamental issues in ultrasonic nondestructive evaluation (NDE) of the various materials and structures, since the effect of flaws on structural integrity is largely dependent on their types. Cracks are usually considered more dangerous than volumetric flaws. Thus, it is common to discriminate only between more crack-like and less volumetric flaws. For weldments, however, it is desirable to introduce more flaw categories such as cracks, porosity and slag inclusions not only for the evaluation of structural integrity of the weld but also for the improvement of weld process performance.
Most of the intelligent ultrasonic flaw classification softwares proposed until now- are turned out to be very sensitive to the operational variables of ultrasonic testing such as transducer characteristics and pulser/receiver settings. This is the very reason for the application of ultrasonic pattern recognition approaches to be still quite rare in realistic field inspections even after the appearance of these softwares.
In this study, we has constructed a database which contains 760 waveforms from crack, porosity and slag inclusion and proposed the "normalized features" which are very robust to the training set.
Normalized features propose in this study demonstrates their capability to reduce the effect of the operational variables not only on the ultrasonic features but also on the bias in the classification due to the variation in the training and the test sets. Thus we feel that using these normalized features an invariant ultrasonic pattern recognition can be realized very efficiently.

목차 (Table of Contents)

• 목차 = ⅰ
• List of Figures = ⅲ
• List of Tables = ⅳ
• ABSTRACT = ⅴ
• 제1장 서론 = 1
• 제2장 이론적 배경 = 6
• 2.1 초음파 형상인식 기법 (Ultrasonic Pattern Recognition) = 6
• 2.1.1 특징추출과 선택 (Feature Extraction and Selection) = 7
• 2.1.2 결함분류 알고리즘 (Classification Algorithms) = 8
• 2.1.3 불변 형상인식 (Invariant Pattern Recognition) = 10
• 2.2 학습근거 분류기 (Learning based classifier) = 12
• 2.2.1 신경회로망(Artificial neural networks) 개요 = 12
• 2.2.2 역전파 신경망 (Back propagation neural networks) = 15
• 2.2.3 확률 신경망 (Probabilistic neural networks) = 18
• 제3장 초음파 결함신호 데이터베이스 구축 = 22
• 3.1 용접내재 결함시험편 = 22
• 3.2 결함신호의 획득 장치 (Acquisition System of Flaw Signal) = 26
• 3.3 제1차 데이터베이스 구축 (Construction of The 1st Database) = 27
• 3.4 제2차 데이터베이스 구축 (Construction of The 2nd Database) = 28
• 제4장 불변 초음파 형상인식 알고리즘 = 42
• 4.1 초음파 결함분류 알고리즘 (Ultrasonic Flaw Classification Algorithm) = 42
• 4.2 특징의 정규화 (Normalization of Features) = 45
• 제5장 결과 및 고찰 = 48
• 5.1 정규화된 특징의 확률밀도함수 (Probability Density Function of Normalized Feature) = 48
• 5.2 불변 초음파 형상인식 알고리즘의 성능 (Performance of Invariant Ultrasonic Pattern Recognition Algorithm) = 50
• 5.2.1 제1차 데이터베이스의 분류성능 (Performance of The 1st Database) = 50
• 5.2.2 제2차 데이터베이스의 분류성능 (Performance of The 2nd Database) = 54
• 제6장 결론 = 58
• 참고문헌 = 59