[4A5] Automatic defect detection and localisation in austenitic stainless steel cladding using ultrasonic inspection
Xuening Zou1,2, Channa Nageswaran1 and Yau Yau Tse2
1TWI Ltd, UK
2Loughborough University, UK
Defect detection in austenitic stainless steel welds and cladding using ultrasound has been an issue for many years. Their inhomogeneous and coarse grain structure causes ultrasound attenuation and scattering. The amplitude resolution is low and defect misinterpretation could happen due to sound skewing. Machine learning (ML) is being explored to help defect detection and identification to overcome those problems. This presentation concerns defect detection and classification using ML approach, specifically convolution neural network (CNN). The traditional ML approach typically requires feature extraction of acquired signal, which is difficult to obtain in the case austenitic stainless steel welded material. The advantage of using CNN is that it can detect the defects without feature extraction.
The dataset for training CNN is derived from finite element simulation as well as limited amount from experiment for validation and testing. Electron backscatter diffraction (EBSD) was used to represent the material property and structure in the model so that it simulates as closely as possible actual sound propagation in austenitic stainless steel welded material. Imaging techniques using full matrix capture (FMC) data is used. FMC allows to capture every possible transmit-receive combination for a given ultrasonic array transducer and therefore provides more information for an effective ML technique.
The dataset for training CNN is derived from finite element simulation as well as limited amount from experiment for validation and testing. Electron backscatter diffraction (EBSD) was used to represent the material property and structure in the model so that it simulates as closely as possible actual sound propagation in austenitic stainless steel welded material. Imaging techniques using full matrix capture (FMC) data is used. FMC allows to capture every possible transmit-receive combination for a given ultrasonic array transducer and therefore provides more information for an effective ML technique.
