Classification performance improvement in imbalanced circumferential guided wave detection data

Abstract

Deep learning has significantly advanced the field of pipeline circumferential guided wave testing. However, the scarcity of defective samples in non-destructive testing datasets poses a significant challenge due to data imbalance. Traditional computer vision augmentation techniques, such as rotation and flipping, are often not directly applicable to guided wave data due to its unique characteristics. To address this issue, we propose a novel approach that combines the strengths of convolutional neural networks (CNNs) and random forest (RF) classifiers. Raw detection data are preprocessed using continuous wavelet transform to generate informative time-frequency images. Random forests enhance predictive accuracy by combining multiple decision trees. The experiment results demonstrate the effectiveness of the proposed CNN-RF approach, particularly in handling imbalanced datasets. When the data imbalance ratio is more than 1/2, the shallow, lightweight CNN could effectively extract signal features and lead to approximately a 56.5% speedup. When the imbalance ratio is less than 1/2, the ResNet34-RF model significantly outperforms the standard CNN model. This approach holds potential for contributing to the miniaturization, lightweight design, and mobility of guided wave detection systems and devices.