基于GRU改进RNN神经网络的飞机燃油流量预测

利用从飞机快速存储记录器(quick access recorder, QAR)中获取的大量数据设计研究了一种利用循环神经网络(recurrent neural network, RNN)及其改进网络门控循环单元(gate recurrent unit, GRU)进行飞机燃油流量预测的模型。首先使用基于时间的反向传播算法(back propagation trough time, BPTT)训练网络,Adam优化算法加速迭代更新神经网络权重。在参数调整实验中发现循环神经网络对历史信息利用能力不足,极易发生梯度消失与梯度爆炸,遂提出改进网络结构,引入GRU重构燃油流量预测模型。在最优的超参数条件下,重构模型在训练集和测试集上的损失函数均方误差(mean squared error, MSE)值分别为0.001 08、0.000 97。通过与朴素RNN的预测曲线和MSE对比可以发现,改进后的GRU网络能够"记忆"更多历史信息而不易出现梯度消失或梯度爆炸的问题,预测精度与曲线拟合能力显著提高。因此,GRU重构模型显著改善了预测能力,并通过实际案例验证该预测模型在故障诊断等领域的应用。

Research on Aircraft Fuel Flow Based on Recurrent Neural Network

A fuel flow prediction model based on RNN (recurrent neural network) and its improved network GRU (gate recurrent unit) is designed and studied based on a large amount of data obtained from aircraft QAR (quick access recorder). Firstly, BPTT algorithm is used to train the network, and Adam optimization algorithm is used to speed up the iterative updating of neural network weights. In the experiment of parameter adjustment, it is found that the cyclic neural network can not make full use of historical information and is prone to gradient disappearance and gradient explosion. Under the optimal hyper parametric conditions, the MSE (mean squared error) values of the reconstruction model on the training set and the test set are 0.00108 and 0.00097 respectively. Compared with the prediction curve and MSE of naive RNN, it can be found that the improved GRU network can "memorize" more historical information without the problems of gradient disappearing or gradient explosion, and the prediction accuracy and curve fitting ability are significantly improved. Therefore, the GRU reconstruction model significantly improves the prediction ability, and the application of the prediction model in fault diagnosis and other fields is verified by practical cases.