TAF2-Net: Triple-Branch Attentive Feature Fusion Network for Ultrasonic Flaw Detection

Abstract

Automatic defect detection is a critical task in the industrial production process. At present, many detection methods based on deep learning have been successfully applied in industrial defect detection systems. However, due to the complexity of the specimen defects, the data collected in practice are imbalanced and have intraclass differences and interclass similarities. These problems may reduce the accuracy of defect detection. This article proposes an ultrasonic defect detection system, which uses a novel triple-branch attentive feature fusion network (TAF2-Net) architecture to detect and classify the defects of specimens. We introduce the multiscale feature extraction (MSFE) module to extract multiscale features from the 1-D ultrasonic echo signal and 2-D time–frequency diagram in two branches and design the deep multimodal feature fusion (DMF2) module to enhance the features at different abstract levels. The fused features are unified through a sequential feature pyramid network (SFPN) and fed into gated recurrent unit (GRU) to extract global sequential features in the third branch. Finally, the features extracted from the multiscale feature extract-1-D (MSFE-1-D) module, multiscale feature extract-2-D (MSFE-2-D) module, and GRU are fused to obtain the final prediction results. Experimental results show that the TAF2-Net is superior to the most advanced method, with the accuracy of 96.54%. Meanwhile, to meet the requirement of hardware systems that lack of computing resources, we also propose a lightweight TAF2-small Net that can still achieve a classification accuracy of 93.6%.