Ultrasonic nondestructive evaluation of solid plates with time reversal and deep neural network

다국어 초록 (Multilingual Abstract)

The Investigation of solid plate using ultrasonic method is common and widely applied in industry field. Plate and plate like structures play an important role in aerospace engineering, nuclear power plant system, submarine and auto industries. Lamb waves, due to its long range propagation and low attenuation energy loss in plate or pipeline structure, in recent years, much attention has been paid to the use of Lamb waves for structure health monitoring. However, the dispersive nature of the Lamb wave in plate structure will cause multi-modes occur, which the wave speed of the Lamb wave is a function of frequency and the pulse shape changes from point to point. When one mode interacts with the defect in plates, mode conversion will also make the received signal complicated. This results in careful excitation of the Lamb wave mode and complex signal processing technique.
In this study, the time reversal technique have been applied to Lamb wave signal processing. According to the linear acoustic wave propagation theory, an input signal can be reconstructed at the excitation point if an output signal is time reversed in time domain and resent at the receiving point. The original input can be well reproduced if the wave propagation in a linear elastic medium, however the discontinuities or defect inside the structure will break the time reversal process. Lamb wave in plates has a variety of modes, time reverse technique can make the dispersed signals to be compressed to their original wave shape. In our study, the FE simulation on Al plate and water medium has proved that the time reversal theory can successfully reproduce the original input signal. However, in real experiment setup, the reconstruction process cannot be achieved because of some time-irreversible elements such as the transducer, the wave function generator etc. Then, the system factor is calculated and a correction factor is proposed to compensate the experiment results. By using the correction factor, the experiment result can also reconstruct the original input.
In the second part of this research, deep neural network is used to classify a variety of Lamb wave FE signals. The low mode antisymmetric Lamb mode is generated at one side and recorded at the other side, in a solid plate with different length and thickness. In industrial applications, the ultrasonic signals are not noise free. The white Gaussian noise is added to the FE simulation signals to augment the database. Without extracting the features, the applied deep neural network to the ultrasonic noisy signals shows a good performance to classify the defect one and non-defect ones.

목차 (Table of Contents)

• CHAPTER 1. INTRODUCTION 1
• 1.1 GENERAL INTRODUCTION 1
• 1.2 REVIEW OF PAST WORK 3
• 1.2.1 Nondestructive testing using Rayleigh and Lamb waves 3
• 1.2.2 Guided waves for structure health monitoring 5
• 1.2.3 Time reversal acoustics 6
• 1.2.4 Deep neural network 7
• 1.3 OBJECTIVES 10
• 1.4 DISSERTATION OUTLINE 11
• CHAPTER 2. FUNDAMENTAL BACKGROUND 13
• 2.1 LAMB WAVES 13
• 2.2 HORIZONTAL SHEAR 15
• 2.3 ULTRASONIC INSPECTION METHOD 18
• 2.4 TIME REVERSAL ACOUSTICS OF LAMB WAVE 19
• CHAPTER 3. GUIDED WAVE TIME REVERSAL PROCESS 24
• 3.1 CONTACT TESTING OF AL PLATE WITH EMAT 24
• 3.1.1 Shear horizontal wave generation mechanism 24
• 3.1.2 Shear horizontal experiment 26
• 3.2 TIME REVERSAL REFERENCE EXPERIMENT 32
• 3.2.1 FE validation of the time reversal recognition in water 33
• 3.2.2 Reference experiment in water 35
• 3.2.3 System function 38
• 3.2.4 Correction factor 39
• 3.3 IMMERSION TIME REVERSAL EXPERIMENT WITH AL PLATE 42
• 3.3.1 Determine the optimum distance for Lamb wave generation 42
• 3.3.2 Immersion testing experiment with Al plate 47
• 3.3.3 Results validation and discussion 52
• 3.4 SUMMARY 55
• CHAPTER 4. DEEP NEURAL NETWORK FOR ULTRASONIC FAW CLASSIFICATION IN PLATES 56
• 4.1 FE VALIDATION OF TIME REVESAL RECONSTRUCTION IN AL PLATE 57
• 4.1.1 Excitation of the tune-burst A0 mode in Al plate 58
• 4.1.2 Reconstruction of input signal 58
• 4.2 ULTRASONIC FLAW DATABASE 64
• 4.3 DEEP NEURAL NETWORK 67
• 4.3.1 Single hidden layer neural network (SNN) architecture 67
• 4.3.2 Deep neural network (DNN) architecture 68
• 4.3.3 Performance of SNN and DNN 68
• 4.4 SUMMARY 70
• CHAPTER 5. CONCLUSION AND FUTURE WORK 71
• 5.1 INTRODUCTION 71
• 5.2 CONCLUSIONS 72
• 5.3 FUTURE WORK 73
• REFERENCES 75