基于Online-RBFNN的航空发动机动态模型研究

Research on aeroengine dynamic model based on Online-RBFNN

为解决航空发动机在出现性能退化时模型精度下降的问题,提出了一种基于在线径向基函数神经网络(online radial basis function neural network,Online-RBFNN)的航空发动机动态模型。采用连续K均值(K-Means)算法和FTRL(follow the regularized leader)在线学习算法,对典型RBFNN进行改进,实现在线学习功能。以某型涡扇发动机正常退化数据为原始样本,建立低压涡轮机(low pressure turbine,LPT)出口总温度动态模型,并与其他多种算法建立的模型进行对比,动态模型的平均绝对误差、均方根误差和校正决定系数分别为0.59、1.7和0.9978;将所建立的动态模型在同型号但不同飞行包线区域、不同退化形式的发动机运行数据上进行测试,模型输出结果的误差可分别控制在[-9,8]K和[-10,9]K范围内。研究结果表明,基于Online-RBFNN的动态模型能有效避免模型精度下降的问题,且具有良好的自适应能力。

In order to solve the problem of model accuracy decline when aeroengine performance degradation occurs,a dynamic model of aeroengine based on online radial basis function neural network(Online-RBFNN)is presented.To realize the online learning function,the typical RBFNN is optimized by using the sequential K-Means algorithm and the FTRL(follow the regularized leader)online learning algorithm.The normal degradation data of a certain type of turbofan engine is used as the original sample to establish a dynamic model of total temperature at low pressure turbine(LPT)outlet.And the dynamic model is compared with other algorithms,while the mean absolute error,root mean square error and adjustment determination coefficient of dynamic model are 0.59,1.7 and 0.9978,respectively.Moreover,a test is performed on the dynamic model by the engine data of the same model with different flight envelope area and degradation forms,the error can be controlled within[-9,8]K and[-10,9]K.The results show that the dynamic model based on Online-RBFNN effectively avoids the problem of model accuracy decline,and has good adaptive ability.