基于深度学习的辅助动力装置性能参数预测方法研究

针对传统机器学习的辅助动力装置(Auxiliary Power Unit, APU)性能参数预测方法不能充分利用参数数据间的时序性和非线性问题，提出了一种基于卷积神经网络(Convolutional Neural Network, CNN)-长短期记忆(Long Short-Term Memory, LSTM)-注意力(Attention)的APU性能参数预测方法。首先，引入一维CNN，通过预处理的参数数据得到不同属性的抽象特征。然后，使用LSTM神经网络对这些特征进行记忆，并结合可以对特征状态赋予不同权重的Attention机制来实现参数预测。使用某型APU的参数数据预测未来不同步长的排气温度(Exhaust Gas Temperature, EGT)。实验结果表明，对于单步EGT的预测，CNN-LSTM-Attention模型在平均绝对百分比误差(Mean Absolute Percentage Error, MAPE)指标上比CNN-LSTM、LSTM和简单循环神经网络(Simple Recurrent Neural Network, Simple RNN)模型分别降低了15.2%、32.5%、60.3%，在均方根误差(Root Mean Square Error, RMSE)指标上分别降低了7.3%、11.6%、32.9%。同时它在多步EGT的预测中具有较高的预测精度，证明了该方法的有效性，为短期APU性能变化趋势预测提供一定的参考。

Prediction Method of Auxiliary Power Unit Performance Parameter Based on Deep Learning

Aiming at the performance parameter prediction method of Auxiliary Power Unit(APU) based on traditional machine learning cannot make full use of the time series and non-linear characteristics between parameter data, a prediction method of APU performance parameters based on Convolutional Neural Network (CNN)-Long Short-Term Memory(LSTM)-Attention is proposed. First, a one-dimensional CNN was introduced, and abstract features of different attributes were obtained through the preprocessed parameter data. Then, LSTM neural network was used to memorize these features and combined with an Attention mechanism that could assign different weights to the feature states to achieve parameter prediction. The parameter data of a certain type of APU was used to predict Exhaust Gas Temperature(EGT) at different steps in the future. The experimental results show that for the prediction of single-step EGT, CNN-LSTM-Attention model reduces Mean Absolute Percentage Error(MAPE) index respectively by 15.2%, 32.5%, and 60.3%, compared with the CNN-LSTM, LSTM, and Simple Recurrent Neural Network(Simple RNN) models, and reduces Root Mean Square Error(RMSE) index by 7.3%, 11.6%, and 32.9%. At the same time, it has higher prediction accuracy in multi-step EGT prediction, which proves the effectiveness of this method and provides a certain reference for short-term APU performance change trend prediction.