Research on Key Technologies for Intelligent Detection of High-Speed Railway Pantograph System Status Based on Deep learning

Abstract: This research proposes an innovative intelligent detection methodology tailored for the high-speed train catenary system, leveraging FPGA-accelerated MobileNetV2. Exploiting the exceptional computational capabilities of the MobileNetV2 convolutional neural network, the methodology incorporates Quantization Aware Training (QAT) to judiciously compress the comprehensive network parameters to one-fourth of the original configuration, ensuring judicious and efficient intelligent detection for the high-speed train catenary system. Notably, the entirety of network weights is strategically allocated to the on-chip resources of the FPGA, effectively circumventing constraints inherent to off-chip storage bandwidth. This strategic allocation addresses power consumption challenges linked to accessing off-chip storage resources, culminating in a substantial augmentation of the real-time operational efficiency of the network.The proposed system, an intricately tuned and energy-efficient Lightweight Convolutional Neural Network (MobileNetV2) recognition system, is meticulously implemented on the Xilinx Virtex-7 VC707 development board. Operating seamlessly at a clock frequency of 200Hz, the system attains an impressive throughput of 170.06 GOP/s with a mere power consumption of 6.13W. The resultant energy efficiency ratio excels at 27.74 GOP/s/W, significantly outpacing the CPU by a factor of 92 and the GPU by a factor of 25. These findings underscore substantial performance advantages when juxtaposed with alternative implementations.

Keywords: high-speed railway; pantograph network monitoring; MobileNetV2;FPGA; deep learning