动态可重构神经网络的FPGA实现及其在燃爆状态监测中的应用

Explosion state monitoring based on FPGA dynamic reconfigurable neural network

提出了一种基于现场可编程门阵列(FPGA),实现动态可重构神经网络的可燃气体燃爆状态监测的方法。传感器采集可燃气体燃爆过程中引起的多种气体浓度变化的数据,在MATLAB中进行离线神经网络训练,燃爆状态的分类预测的准确性依赖于已训练好的模型。基于片上网络架构在FPGA上实现动态可重构神经网络,并优化了其硬件设计,其能够现场切换、调整合适的网络模型对气体燃爆状态进行在线的分类预测,以达到最佳预测效果。实验结果表明,文中所提出的方法对气体燃爆的多种状态的分类预测具有较高的准确率,该方法提高了硬件化神经网络在燃爆状态监测应用中的灵活性、通用性及分类准确性。

This paper proposes a method for monitoring the explosion state of combustible gas based on Field-Programmable Gate Array(FPGA)dynamic reconfigurable neural network. The data of various gas collecting by sensors are used for offline-training in MATLAB. The accuracy of classification of explosion state depends on the trained model. Based on the Network on Chip architecture,the optimized dynamic reconfigurable neural network is able to configure appropriate neural network model to predict the explosion state online. The achieved experimental results seem to show the effectiveness of the explosion state monitoring system. It improves the flexibility and classification accuracy of the hardware implementations of neural network.