Real-time rail surface defect detection: a neuromorphic approach

Abstract

This study explores the implementation of a neuromorphic-based approach for real-time rail surface defect detection, addressing the growing demand for accurate and efficient maintenance solutions in rail systems. Leveraging a dataset of 1514 rail surface images collected under real-world operating conditions and annotated with binary labels ("defect" and "no defect"), the research applies a comprehensive data augmentation pipeline - including rotation, flipping, zooming, and contrast adjustments - to expand the dataset to 248374 images. This augmented dataset ensures class balance and greater variability, enhancing model generalisation and preventing overfitting. A consistent data split (70% training, 15% validation, 15% testing) is employed to maintain fairness across evaluations. The study presents a comparative evaluation of spiking neural networks (SNNs) and convolutional neural networks (CNNs) in terms of classification performance, energy efficiency, and real-time feasibility across three hardware platforms: a 12-core AMD Ryzen 9 CPU (3.8 GHz), a NVIDIA RTX3060 GPU (12 GB VRAM), and the BrainChip AKD100 neuromorphic processor. The analysis further investigates the influence of hand-crafted features - including Grey-Level Co-occurrence Matrix (GLCM), Local Binary Pattern (LBP), and Sobel and Canny edge detection - on model performance and computational efficiency across different hardware configurations. Experimental results show that CNNs consistently outperform SNNs in precision and recall, particularly when implemented on GPUs, where the CNN model achieves a precision of 94.82%, recall of 94.88%, and F1-score of 94.49%. The Akida neuromorphic processor using a CNN model, with its 8 neural cores and sub-1W power consumption, maintains a high precision of 94.74% while achieving latency below 5 ms and throughput exceeding 250 frames per second - demonstrating strong balance of accuracy, speed, and energy efficiency. In contrast, SNNs, though more energyefficient, show diminished accuracy, with the CPU-based SNN model achieving an F1-score of only 83.75%. The research also outlines optimal hardware configurations for deployment, considering factors such as sensor quality, data acquisition methods, power supply, and cooling requirements. The findings indicate that combining CNNs with the Akida processor offers an ideal solution for real-time rail defect detection, balancing classification performance with resource efficiency. The final chapter synthesises these insights to offer practical guidance for system deployment and highlights future directions to advance neuromorphic-based approaches in railway maintenance.