基于深度学习OFDM信道补偿技术硬件实现

为了解决部分高性能深度学习神经网络因存在复杂度高及计算量大等缺陷在嵌入式设备中应用效果不理想的问题；以小型化集成智能无线电设备AIR-T为平台实现了基于深度学习的OFDM信道补偿技术；在FPGA芯片上不仅实现了OFDM信号传输系统模块,也实现了传统信道估计与均衡模块,模块对数据进行预处理减轻神经网络工作量以完成神经网络信道补偿技术模块在Jetson TX2平台GPU上的高效实现；由实验记录神经网络训练过程中的计算复杂度和参数拟合速度得知,传统信道估计与均衡模块有效降低了网络训练时的运算次数；由测试性能方面可知,经过神经网络信道补偿后的数据误码率比之前传统信道估计与均衡后的误码率有明显降低；

Hardware Implementation of Deep Learning Based OFDM Channel Compensation Technique

In order to solve the problem that some high-performance deep learning neural networks are not ideal for application in embedded devices due to the defects of high complexity and large computation. Deep learning-based Orthogonal Frequency Division Multiplexing(OFDM) channel compensation technology is implemented on the Artificial Intelligence Radio-Transceiver(AIR-T), a miniaturized integrated smart radio device, as a platform. Not only the OFDM signal transmission system module, but also the conventional channel estimation and equalization module are implemented on the Field Programmable Gate Array(FPGA) chip.These modules preprocesses the data to reduce the workload of the neural network in order to complete the efficient implementation of the neural network channel compensation technology module on the graphics processing unit(GPU) of Jetson TX2 platform. The computational complexity and parameter fitting speed of the neural network training process are recorded, and the conventional channel estimation and equalization module effectively reduces the number of operations during the network training. From the tested performance aspects, it can be seen that the data BER after the neural network channel compensation is significantly lower than the BER after the previous conventional channel estimation and equalization.