长短期记忆网络驱动的脉冲感应式推力器放电特征研究

Discharge Characteristics of Plasma Inductive Thruster Driven by Long Short-Term Memory Network

关键词:  脉冲感应式推力器  放电电路模型  长短期记忆网络  放电序列数据  采样预测

Key words:  Plasma inductive thruster  Discharge circuit model  Long short-term memory network  Discharge sequence data  Sampling and prediction

摘要:

为快速与可靠预测脉冲感应式推力器的放电特性，提出了一种改进的等离子体放电电路模型与长短期记忆网络结合（MPDCM-LSMT）的数值实验方法，建立了适用于该型推力器的等离子体放电幅值序列数据生成、训练和采样预测的融合模型。为生成高质量的序列，以等离子体电磁感应和流动方程为核心，发展了Ar，He和N₂的热力学计算方法，并根据不同的能量沉积和电导率模型，推导出三种电路模型。通过计算的冲量和放电曲线与实验对比分析，识别最优模拟推力器放电特性的数据模型。对电压和电流数据集训练并采样后发现，质量恒定的情况下，采用融合模型训练21组序列数据得到的LSMT网络，可实现主放电阶段趋势的预测。在文中研究的范围内，对高初始放电电压条件下主放电周期的预测发现，电压曲线与计算曲线吻合度高，电流曲线峰值误差小于3.8%，对应时间误差小于5.3%。结果表明，实现推力器放电预测所需的网络层数和单元数与样本量密切相关，层数影响放电变化趋势预测正确性，单元数则影响曲线的平滑程度。

Abstract:

To predict discharge characteristics of the Pulsed Inductive Thruster （PIT） fast and reliably， this work focuses on a numerical experiment method that combines the Modified Plasma Discharge Circuit Model with the Long Short-Term Memory （MPDCM-LSMT）， and builds a merging model capable of generating， training and sampling discharge sequential data. To generate high-quality sequential data， three circuit models are developed based on different energy depositions and electrical conductivities. The models put electromagnetic and flow equations of plasma as their hearts， and incorporate thermodynamic models for Ar， He and N₂. And the best data model is identified by comparing the impulses and discharge curves calculated with the experiments. By training and sampling voltage and current data sets using the merging model， the work suggests that the LSMT model owning 21 discharge curves will reasonably predict the trend of the main pulse when the propellant mass is determined. In the main pulse under high initial discharge voltages within the range studied in this work， the predicted voltage curves coincide well with those calculated by the data generation model， while for the current curves， relative error of the peak is less than 3.8% and the corresponding time error is within 5.3%. The results show the number of layers and units of the network strongly relates to the samples， and the number of layers affects the predicted trend， while the number of units determines smoothness of the curves.