Ma, Sean LongyuFu, Yulin2025-03-102025-03-102025https://hdl.handle.net/2292/71608Railway fault detection plays a crucial role in ensuring the safety and reliability of the railway infrastructure. In this study, we propose a real-time railway fault detection system implemented on Field-Programmable Gate Arrays (FPGAs), leveraging the efficiency of Binary Neural Networks (BNNs) for edge computing applications. BNNs, characterized by 1-bit weights and activations, offer significant advantages for low-latency, real-time monitoring tasks due to their efficient bitwise operations and minimal memory requirements. This makes them particularly well-suited for embedded systems, where low power consumption, high throughput, and real-time decision-making are essential. To address the challenge of resource utilization imbalance in FPGA-based systems, we introduce a novel DSP packing algorithm specifically designed to optimize Digital Signal Processor (DSP) usage while minimizing reliance on Lookup Table (LUT) resources, two primary resources in FPGA architecture. This approach effectively balances the usage of FPGA resources, enabling efficient deployment of the model on resource-constrained platforms. Experimental results show that our system achieves state-of-the-art detection accuracies of 90.9% and 93.4% with the RBPnet2 and RBPnet4 models, respectively. Furthermore, our FPGA-based solution outperforms conventional Central Processing Units (CPUs) and Graphics Processing Units (GPUs), achieving 82.77× and 9.64× improvements in energy efficiency, respectively. This work highlights the significant potential of FPGA-based BNNs for real-time railway fault detection, offering a robust and energy-efficient solution for edge computing applications. The high performance and low power consumption of the system demonstrate its suitability for deployment in consumer electronics and industrial IoT systems, where real-time monitoring and localized data processing are critical.https://researchspace.auckland.ac.nz/docs/uoa-docs/rights.htmBNNCNNFPGAEdge AI SystemAcceleratorThesisCopyright: the authorAttribution-NonCommercial-NoDerivatives 4.0 Internationalhttps://creativecommons.org/licenses/by-nc-nd/4.0/