基于递归径向基神经网络的航空发动机盘腔瞬态壁温预测

Transient Wall Temperature Prediction of Aero-Engine Cavity based on Recurrent Radial Basis Function Neural Network

关键词:  航空发动机  空气系统  瞬态壁温预测  RBF神经网络  多维重构

Key words:  Aero-engine  Air system  Transient wall temperature prediction  RBF neural network  Multi-dimensional reconstruction

摘要:

针对航空发动机空气系统盘腔瞬态壁温动态预测难的问题，提出了一种基于径向基神经网络的递归预测模型。通过时序数据多维重构的方法建立训练样本，强化径向基神经网络对“时滞性”的预测能力，分析了模型固有超参数和由多维重构引入抽样控制参数对模型预测精度的影响。采用简化的典型盘腔壁面换热模型结合公开的试验历程转速数据，构建了供模型训练和测试的瞬态壁温数据样本，以递归调用模型的方式完成了对测试样本时序数据的预测和验证。结果表明，与常规的径向基神经网络预测模型相比，该模型的平均相对预测偏差由3.0%降低至0.45%，有效提升了模型的预测精度。为航空发动机盘腔瞬态壁温异常监控及超温排故问题提供了一种新的预判方法。

Abstract:

Aimed at the problem that the dynamic prediction is difficult for transient wall temperature of aero-engine air system cavity， a recurrent prediction model based on Radial basis function（RBF） neural network is proposed. Firstly， the training sample is constructed through the method of multi-dimensional reconstruction of time series data to enhance the prediction ability of RBF neural network for time lag characteristic. The influences of the inherent hyperparameters and the sampling control parameters introduced by multi-dimensional reconstruction on the prediction accuracy of the model are analyzed. Finally， the transient wall temperature data samples for model training and testing are constructed by using a simplified typical wall heat transfer model combined with publicly available experiment process rotate speed data. And the prediction and validation for the testing data are accomplished via the recursive invoking model mode. The results show that the average relative prediction deviation of the proposed model is reduced from 3% to 0.45% in comparison with the conventional RBF neural network prediction model， which effectively improves the prediction accuracy of the model. This paper provides a new prediction method for the monitoring of transient wall temperature anomalies and over-temperature exclusion of aero-engine cavity.