Evaluation of subsurface defects in metallic structures using laser ultrasonic technique and genetic algorithm-back propagation neural network
| Autor: | Yan Jun Liu, Kaixing Zhang, Shifeng Guo, Wei Feng, Chen Dan, Gaolong Lv |
|---|---|
| Rok vydání: | 2020 |
| Předmět: |
010302 applied physics
Materials science Artificial neural network business.industry Mechanical Engineering Acoustics Condensed Matter Physics Laser 01 natural sciences law.invention Back propagation neural network symbols.namesake law Nondestructive testing 0103 physical sciences Genetic algorithm symbols General Materials Science Ultrasonic sensor Rayleigh scattering business 010301 acoustics Reliability (statistics) |
| Zdroj: | NDT & E International. 116:102339 |
| ISSN: | 0963-8695 |
| Popis: | An effective nondestructive evaluation technique that enables the detection and quantification of subsurface defects is highly demanded for assuring safety and reliability of safety-critical structures. In this work, an improved genetic algorithm-back propagation neural network (GA-BPNN) model and non-contact laser ultrasonic technique are combined to quantify the width of subsurface defects. An experimentally validated numerical model that simulates the interaction of laser-generated Rayleigh ultrasonic waves with subsurface defects is firstly established, which is further utilized to generate a large number of labeled laser ultrasonic signals for training the GA-BPNN model. A total number of 189 data are obtained from simulation and experiments, with 173 simulated signals for training the GA-BPNN model and the remaining 13 simulated signals together with 3 experimental signals for verifying the performance of the trained GA-BPNN model. Five features including three time-domain features (maximum, minimum and peak-to-peak value of the Rayleigh ultrasonic waves) and two frequency-domain features (Fc, BW-6dB), which are identified sensitive to the width of subsurface defects by both experiments and simulation, are extracted as inputs to train the machine learning algorithm. The result demonstrates that the GA-BPNN model trained with the combination of time and frequency features has the average error of 2.15%, which is substantially smaller than the errors obtained from the model trained with only time-domain features and frequency-domain features, with the average errors of 4.43% and 21.81%, respectively. This work proves the feasibility and reliability to quantify the width of subsurface defects in metallic structures using laser ultrasonic technique and the improved GA-BPNN algorithm. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|
Full Text from ScienceDirect