Reinforcement Learning Based Neural Architecture Search for Flaw Detection in Intelligent Ultrasonic Imaging NDE System
Xin Zhang, Jafar Saniie
Abstract
Ultrasonic flaw detection has been extensively used for NDE applications because it has high inspection resolution and accuracy. Conventional ultrasonic flaw detection is more vulnerable to human errors and time-consuming as the workload increases. The artificial intelligence (AI), such as machine learning (ML) methods, automates the evaluation process and is more reliable and practical. However, modeling the ML algorithms, such as the neural networks (NN) requires substantial computational resources for training and significant effort in obtaining efficient NN architecture. In this study, we introduce a reinforcement learning (RL) based neural architecture search (NAS) framework to automatically model the optimal NN design. By using this framework, a NAS-based NN: Ultrasonic Flaws Detection NAS Neural Network: UFDNASNet, is proposed for flaws detection with high accuracy and data-efficiency. The ultrasonic datasets are processed by the NAS framework using the recurrent neural network (RNN) controller to search for the best convolutional operations. The flaw detection performance is analyzed and compared between the introduced UFDNASNet and several hand-designed deep Convolutional Neural Networks (deep-CNN) based on detection accuracy and inference data-efficiency. To evaluate the performance for defects detection, the NNs are trained with the transfer learning (TL) using the USimgAIST dataset of B-scan images representing without-defect and with-defects cases. The B-scan images were collected by using the pulsed laser ultrasonic scanning system from 17 stainless steel specimen plates with various types of flaws and some plates without any damage. Our purpose is to realize an intelligent system to detect flaws with high accuracy for data-efficient ultrasonic NDE applications.