Improving Generalization Performance by Data Augmentation Using SpecAugment and GAN for Concrete Impact-Echo Classification

Abstract

Automation of impact-echo monitoring, one of the non-destructive inspection methods for concrete structures, has been studied, including quantification of impact-echo features by signal processing and discrimination by machine learning. For automation, it is necessary to perf orm machine learning based on labeled data prepared in advance, and only prediction should be performed on the test site using the pretrained model. However, the accuracy of discrimination generally decreases when data is collected at different locations. In this study, we apply conditional adversarial neural network (CGAN) and data augmentation by SpecAugment to methods that use convolutional neural networks (CNNs) to discriminate images that have the time-frequency characteristics of impact-echo and attempt to improve generalization performance for test data from different sampling locations. We show that data augmentation by SpecAugment is effective in improving generalization performance.