Abstract: In order to achieve accurate flow characteristics model of the disc type control valve and improve the accuracy of inlet flow prediction of altitude cabin, modeling method based on BP neural network and NARX network is proposed. Based on the analysis of the position of the control valve and sensor point, the control valve and the pipe behind the valve are modeled as a whole. The modeling methods of the flow coefficient, static BP neural network, and dynamic NARX network based on Gamma Test (GT), are compared, and suggestions on selecting in engineering are given. Finally, based on the test flow data, the steady-state error and dynamic error of the flow output of each model are simulated and compared when the valve changes differently. The results show that the modeling precision of the BP neural network method and the NARX network method is better than the flow coefficient method. Meanwhile, the maximum steady-state error of the BP neural network model is 0.52kg/s, which is preferable to the models of NARX network and flow coefficient. Compared with the models of BP neural network and flow coefficient, the NARX network model whose maximum dynamic error is 2.04kg/s, can show the dynamic flow characteristics accurately.

Key words: Altitude simulation test facility;Special control valves;Flow characteristics;NARX network;Dynamic model

摘要： 为了获得准确的轮盘式特种调节阀流量特性模型，提高高空舱进口流量预测精度，提出了基于BP神经网络和NARX网络的建模方法。在对调节阀与传感器测点位置分析的基础上，将调节阀和阀后容腔作为整体进行建模。对比研究了流量系数、静态BP神经网络以及基于Gamma Test的动态NARX网络建模方法，并给出了工程中选取建模方法的建议。以试验流量数据为基准，仿真对比了不同阀门开度变化时，各模型输出流量的稳态误差和动态误差。结果表明，BP神经网络方法和NARX网络方法建模精度要优于流量系数法。同时，BP神经网络模型最大稳态误差为0.52kg/s，优于NARX网络模型和流量系数模型。NARX网络模型的最大动态误差为2.04kg/s，相比于BP神经网络模型和流量系数模型，能够更准确地反映流量的动态特性。

关键词: 高空台;特种调节阀;流量特性;NARX网络;动态模型