星载天线知识与数据深度融合的指向精度物理感知状态评估模型构建

The core problem of reducing antenna weight and improving service performance is how to quantitatively assess pointing accuracy degraded state of satellite on-board antenna. Pointing accuracy degradation is a nonlinear natural degradation process govered by multi-factor superposition, multi-component synthesis. As monitoring data are referred to variable working conditions, multiple physical domains and nonlinearity, it is too hard to overcome the quantitative assessment problems by merely concentrating on the perspective of "mechanism knowledge" or "monitoring data". This project focuses on the urgent engineering requirements of quantitatively assessing the pointing accuracy degraded state of satellite on-board antenna. For the problem how to identify the working conditions of the multi-source data with very similar profile, a physical perception deep belief network model with input and output double-layer limited Boltzmann machine is proposed that can deeply combine the condition knowledge with the variable condition data. For the problem that it is difficult to extract the precision degraded feature from the multi-source data with difference physical background in a certain condition, a multi-layer physical perception deep automatic encoder Network model is proposed with depth fusion of multi-physics knowledge and multi-source monitoring data. For the problem that it is difficult to quantitatively assess the degraded state because of the nonlinear stochastic mapping relationship between degraded features and degraded states, a physical perceptual non-homogeneous hidden semi-Markov model is proposed with the degraded state knowledge and degraded feature data in hidden state layer. The project will effectively solve the quantitative assessment problem of the pointed accuracy degraded state, which can provide scientific basis for reducing antenna weight in development stage and improving the service performance in operation stage.

星载天线指向精度退化状态定量评估问题是减小天线重量、提升服役性能的核心问题。指向精度退化是一个多因素叠加、多部件综合的非线性退化过程，监测数据涉及变工况、多物理域和非线性，仅从机理知识或监测数据角度难以解决定量评估问题。本项目从天线指向精度退化状态定量评估的迫切工程需求出发，针对不同工况多源数据高度相似难以识别工况的问题，提出工况知识与变工况数据深度融合的受限玻尔兹曼机输入输出双层物理感知深度置信网络模型；针对单工况多源数据物理背景差异大难以提取指向精度退化特征的问题，提出多物理场知识与多源监测数据深度融合的多层物理感知深度自动编码器网络模型；针对退化特征与退化状态之间非线性随机映射关系复杂难以定量评估退化状态的问题，提出退化状态知识与退化特征数据深度融合的隐含状态层物理感知非齐次隐半马尔科夫模型。本项目将有效解决指向精度退化状态定量评估难题，可为减小天线重量、提升服役性能提供科学依据。

本项目从天线指向精度退化状态定量评估的迫切工程需求出发，针对不同工况多源数据高度相似难以识别工况的问题，完成了工况变化规律和任务剖面等工况先验知识在深度神经网络模型多层结构中的相容互斥物理感知表达方法研究，结合多源变工况数据类间距离最大算法建立了工况知识与多源变工况数据深度融合的物理感知深度网络模型；针对单工况多源数据物理背景差异大难以提取退化特征的问题，完成了多物理场先验知识在深度自编码器网络模型多层结构中的结构化确定表达方法研究，结合多源数据误差正反向传播算法和参数自动微分方法建立了多物理场先验知识与多源单工况监测数据深度融合的物理感知深度自动编码器网络模型；针对退化特征与退化状态之间非线性随机映射关系复杂难以定量评估退化状态的问题，完成了退化状态的非线性多状态随机变化规律在随机过程模型隐含状态层的物理感知表达方法研究，结合退化特征数据建立非线性分岔退化状态知识与连续退化特征数据深度融合的退化状态物理感知评估模型。利用获取的多源变工况监测数据对建立的工况分类模型、退化特征提取模型和退化状态定量评估模型进行了验证。本项目研究成果可为航天、轨道交通等行业的关键设备研制阶段减轻重量、运维阶段提升服役性能提供科学依据，具有重要的工程价值和科学意义。