A detection method for rail internal defects based on data augmentation and lightweight deep network

Abstract

Detecting internal defects in degrading rails to ensure transport capacity and safety in railway engineering has consistently been a critical concern. However, current ultrasonic-based detection methods exhibit limitations in efficiency and accuracy, especially under limited computational resources. To address this challenge, this paper proposes a method that involves a specialized B-scan image processing pipeline and a light-weight detection model for rail internal defects (LRID). Specifically, the pipeline comprises normalization, multi-channel partition filtering, and data augmentation utilizing traditional transformations and generative artificial intelligence model. The pipeline aims at enhancing signal-to-noise ratio and learnability of the B-scan dataset. For constructing the LRID, a scaled version of real-time detection transformer–large (RTDETR-L) is leveraged as the baseline. The encoder of the baseline is first pruned to reduce complexity. Then, a new convolutional block called ghost-rep-ghost (GRG) is proposed to build the backbone of the LRID, which combines compactness and multi-branch learning. Comparative evaluations of detection performance across various models demonstrate the superiority of the LRID. The model achieves a detection speed of 130 ${\tau }_{FPS}$ and exhibits a significant improvement in ${\theta }_{mAP}$(0.5:0.95) from 64.5% to 68.2% and a decrease in giga floating point operations per second (GFLOPs) from 17.3 to 16.8, relative to the baseline. Ablation experiments further reveal that the backbone contributes a 4.8% improvement in ${\theta }_{mAP}$(0.5:0.95), and elucidate the structural design of the backbone. Ultimately, the detection accuracy for most defect classes within B-scan dataset approaches 90%, indicating the effectiveness and feasibility of the proposed method.