FPGA平台轻量化卷积神经网络辐射源信号识别方法

Emitter Signal Identification Method with Lightweight CNN on FPGA Platform

针对卷积神经网络计算资源消耗大、难以在边缘侧应用等问题,提出了一种面向FPGA(Field Programmable Gate Array)平台的基于知识蒸馏的轻量化卷积神经网络辐射源信号识别方法。该方法以信号时频图作为特征提取对象,结合改进的知识蒸馏方法对卷积神经网络进行轻量化处理,通过注意力图增强知识信息传递,并融合深度可分离卷积,进一步提高网络稀疏度。最后,将该轻量化网络在FPGA平台上进行结构优化,通过改进循环策略和流水线并行设计,加速轻量化卷积神经网络的辐射源信号识别过程。仿真结果显示,利用本文提出的轻量化卷积神经网络辐射源信号识别算法,网络参数量降低了81.8%,在信噪比不低于-12dB的条件下,信号识别准确率达到了90%以上,FPGA平台信号识别时间为86ms,平均功耗为2W,可满足边缘侧终端对信号实时检测以及低功耗的实际应用需求。

To address the problems of large computational resource consumption of convolution neural networks(CNNs)and difficulty in edge-side applications,this paper proposed a method for emitter signal recognition on FPGA(Field Programmable Gate Array)platforms.The method used a lightweight CNN based on knowledge distillation.It took the time-frequency maps of signals as the feature extraction object for CNNs lighten with the improved knowledge distillation method.The attention maps were used to enhance the transfer of knowledge information.Furthermore,the network’s sparsity was improved by fusing depthwise separable convolution neural networks.Finally,the lightweight network was structurally optimized on the FPGA platform,including improving the cyclic strategy and using pipeline,to accelerate the process of signal recognition.The simulation results showed that the parameters was reduced by 81.8%owing to the proposed light-weighted CNN.The recognition accuracy exceeded 90%under the condition that the signal-to-noise ratio is not less than-12 dB.The recognition time is 86ms and the average power consumption is 2W when test on FPGA platform,which can meet the practical requirements of edge-side terminal for real-time signal detection and low power consumption.