An Efficient FPGA-Based Edge AI System for Railway Fault Detection

Binary neural networks (BNNs), characterized by 1 bit-width weights and activations, are well-suited for deployment on edge computing platforms such as field-programmable gate arrays (FPGAs), especially for real-time monitoring and detection tasks due to their efficient bitwise arithmetic and minimal memory requirements. This makes them ideal for energy-sensitive consumer electronics, where edge computing lowers latency, enhances privacy, and conserves bandwidth by enabling real-time decision-making and localized data processing. In this study, we leverage the strengths of BNNs to develop a real-time railway fault detection system on FPGAs. To address the imbalanced resource utilization that limits model implementation in resource-constrained scenarios using FPGAs, we have developed a digital signal processor (DSP) packing algorithm that optimizes DSP usage and reduces reliance on lookup table resources. Our experiments demonstrate that, with optimized FPGA resource utilization, the system achieves state-of-the-art accuracies of 90.9% and 93.4%, respectively. Furthermore, our FPGA-based system outperforms central processing units and graphics processing units, achieving 101.9× and 33.8× improvements in cost-energy efficiency. This advancement in FPGA-based system showcases the potential of edge computing in transforming consumer devices by enabling sophisticated functionalities while maintaining high energy efficiency. We release our source code to encourage further exploration in this direction, and the code is available at: https://github.com/JiuYing2235/BNN_Railway.git.