Automatic segmentation of ultrasonic TFM phased array images: the use of neural networks for defect recognition

Abstract

Ultrasonic Phased Array imaging is a key method for fast and reliable nondestructive testing of structures, especially when only one side of the part is accessible. Full matrix capturing (FMC) in combination with the total focusing method (TFM) provides a strong tool for ultrasonic imaging of structures with complex flaw patterns. However, still, operator needs to go through the generated images and manually check for the possible defects. One important task is to separate true and false indications, as some of them are noises or artifacts. Inspecting large structures with TFM Phased Array Imaging generates a huge amount of data which takes a significant time to go through them manually. In this work, we evaluate the possibility of using the neural network as an artificial intelligent toolbox to identify the defects. Using finite element method and an in-house developed TFM code, the phased array images are produced as the input to the neural network. The output of the neural network, target, is defined as the probability of defect existence. After generating TFM final images with different flaw patterns, the network was trained and evaluated based on the stochastic genetic algorithm learning method. This made the training feasible with limited provided data. Results indicate the great potential of machine learning for automatic or assisted defect recognition. The main challenge to pursuing a comprehensive and reliable machine learning toolbox, is to train the system with a satisfactory number of examples in different situations to ensure the final product is able to cover all possible conditions. It is concluded the proposed neural network model is capable of image pattern recognition with limited provided training data.