Abstract: In order to solve the problem of liquid-propellant rocket engine (LRE) failure prediction accurately，reliably and rapidly，a failure prediction method based on error predicting correction is proposed. The prediction model of wavelet process neural network (WPNN)，established based on historical data，is firstly adopted to make failure prediction of liquid-propellant rocket engine. The prediction error of every sample is recorded for setting up error prediction model，which is a double parallel feedforward discrete input process neural network (DPFDPNN). Then the prediction error is estimated with DPFDPNN model to compensate the predicted value with the WPNN model. Finally，the proposed method is validated through a test case with ground testing data of LRE hydrogen turbopump. The results show that，the proposed method outperformed the WPNN model at accuracy and adaptability. The normalized root mean square error (NRMSE) of the prediction value is 0.392，when the number of prediction step is 10，and the average computing time is 76ms. The proposed method may be used to solve the problem of LRE failure prediction.

Key words: Liquid-propellant rocket engine；Failure prediction；Wavelet process neural networks；Double parallel feedforward discrete input process neural network；Error correlation

摘要： 为解决液体火箭发动机故障预测这一难题，提出一种基于误差预测修正的故障预测方法。在历史数据的基础上建立小波过程神经网络故障预测模型，同步计算学习样本的预测误差，根据上述误差建立双并联离散过程神经网络预测模型。预测时，将预测误差值实时补偿到小波过程神经网络预测模型以提高预测精度。通过液体火箭发动机地面试验中的涡轮泵数据对该方法进了验证。结果表明，该方法在预测精度和适应能力上较单一的过程神经网络预测模型有显著提高，进行10步预测时，预测值的标准化均方根误差为0.392，预测平均耗时为76ms，能够用于解决液体火箭发动机故障预测问题。

关键词: 液体火箭发动机；故障预测；小波过程神经网络；双并联离散过程神经网络；误差修正