【作者】 岑朝辉；

【导师】 魏蛟龙；

【作者基本信息】 华中科技大学 ， 通信与信息系统， 2011， 博士

【摘要】 卫星姿态控制系统是保障卫星正常运转的核心子系统。由于卫星在太空运行过程中具有远程不可及等应用特殊性,其姿态控制系统的故障诊断问题一直备受航天工业界学者及研究人员关注。卫星星上资源和人工干预能力有限,太空监测环境恶劣和不确定性因素多等特点决定了卫星故障诊断不仅要具备一般故障诊断的可靠性及准确性要求,还必须具有快速自诊断及自主容错恢复能力。面向我国重大设施及装备的高可靠性及长寿命战略发展需求,本文在863高技术计划项目(2007AA04Z438)资助下,以卫星姿态控制系统中的实际故障问题为切入点,从模型(定量)途径、智能(定性)途径及模型与智能混合途径三方面系统研究了故障检测、分离、估计与恢复等故障诊断问题及技术。基于对象系统模型研究了卫星姿控系统中的故障诊断与容错问题。针对执行器非完全失效这一典型故障,提出了以自适应观测器与扩张状态观测器相结合作为故障诊断与估计手段、采用四元数控制律故障调节技术作为恢复措施的闭环被动容错策略,可实现执行器非完全失效故障情形下的卫星姿控系统故障检测、分离与恢复。针对传感器故障失效这一问题,提出了一种以奉献观测器作为故障检测与隔离手段、采用KX观测器实现传感器故障容错观测的闭环主动容错方法,可克服因部分传感器观测失效导致的整体观测失效及闭环控制不稳定问题,使得卫星姿态控制系统仍能被部分观测并保障其闭环控制稳定性,实现传感器失效情形下的卫星姿态控制系统故障检测、分离与恢复。基于计算智能研究了以神经网络模式分类为核心方法、面向卫星姿态控制系统实时信号的故障诊断新方法。针对红外地球仪故障,采用双层ELMAN动态神经网络,对故障时域信号进行样本学习及模式分类,实现故障检测与故障隔离。针对故障检测的高实时性要求与组合故障隔离问题,引入改进平移小波实时获取故障信号奇异点的模极大值,避免故障检测依赖样本学习的缺陷,在故障检测基础上采用改进动态循环神经网络(Improved Dynamic Recurrent Neural Network, IDRNN)进行智能故障识别,可以实现面向监测实时信号的故障检测、及复合模式分类。基于模型与计算智能混合途径研究了以神经网络模型辨识为核心方法、面向飞轮进行离线辨识与在线观测结合的故障诊断新方法。针对飞轮正常及故障模式,通过离线设计与训练多个BPNN神经网络辨识模型并作为估计器在线监测生成残差,可实现基于残差的飞轮故障检测及故障程度区分。针对优化模型精度实现准确故障评估的问题,提出了一种灰盒神经网络模型辨识与故障评估方案,通过直接继承对象系统动力学并采用改进的自定义激励训练策略,获得正常模式下的灰盒神经网络辨识模型并作为估计器在线嵌入生成残差,可避免非线性动态系统模型辨识中由于动力学不匹配所导致的模型精度下降缺陷,不依赖故障模型即可实现故障检测及故障评估。研究了验证故障诊断方法及结论的相关基础软硬件平台构建问题。构建了一套基于xPC Target的卫星姿态控制系统实时数学仿真软硬件平台,以某型三轴对地定向卫星为蓝本,建立了用于故障注入及故障模拟的完整卫星姿控系统模型,基于该平台可实现故障征兆分析及实现对上述诊断方法的验证。更多还原

【Abstract】 Satellite Attitude Control System (SACS) is the key subsystem of an artificial satellite. As the satellites are unreachable in remote space, Fault Diagnosis (FD) on SACS is a hot and frontable problem in the FD research area. More importantly, satellite also has other disadvantages, such as the limited onboard resource in satellites, limited operation ability by human being, bad space environment, and more uncertainty. Those entire disadvantages make a requirement that FD of Satellite must has the ability of quickly self-diagnosis and the ability of being tolerant with fault itself. With the requirement on reliability and long-life of key plant or equipment of china and sponsored by china 863 high technique program (2007AA04Z438), aimed on the practical problem of SACS, this thesis do researches on fault detection, isolation, estimation and recovery from three main approach:model-based (quantitive), intelligent-based (qualitative), hybrid-based.Following the model-based approach, fault diagnosis and fault-tolerant problems of SACS are researched. Aimed on the partial loss of effect (LOE) fault of actuators, a combination with adaptive observer and extended state observer is proposed to diagnosis faults and estimate the severity of fault, and fault accommodation based on quaternion-feedback control low is proposed to recover the fault and implement passive fault tolerant in close-loop. Based on the work above, the partial loss of effect (LOE) fault of actuators in SACS can be detected, isolated and accommodated. Aimed on the partial loss of effect (LOE) fault of sensors, FDI based on devoting observer and active fault-tolerant measures with sensor fault in close-loop fault-tolerant based on KX observer is proposed to overcome full LOE and close-loop instability, which results from that partial sensor is wrong, and make SACS being observed in part and stable in close-loop. Based on the work above, Aimed on the partial loss of effect (LOE) fault of sensors in SACS can be detected, isolated and recovery.Following the intelligent-based approach, we do research on FD of SACS based on the real-time signal by using some neural network (NN) techniques. Aimed on the fault of infrared earth sensor, double ELMAN dynamical NNs are used to learn from the transient signal and classify the specified fault modes in order to detect and isolate the faults. Aimed on the real-time requirement of FD and isolation on multiple faults, an improved MALLET wavelet is introduced to obtain the maximum modulus of signal singular point and avoid learning from samples in order to detect fault, and then an improved IDRNN (Improved Dynamic Recurrent Neural Network) is proposed to identify the faults. Based on the works above, we can detect faults and isolate multiple faults based on real-time monitoring signal.Following the hybrid-based approach, a novel FD for Reaction Wheel (RW) by using NN model identification is proposed. Aimed on the fault mode and fault-free mode in RW, multiple BPNN identification model are designed and trained offline, and then all of them are embedded into object system in order to generate diagnosis residuals online. Based on the residuals, RW faults with different severities can be detected and differed. Aimed on the fault estimation (FE) problem depending on model accuracy, grey box neural network model identification and fault estimation is proposed. By inheriting the dynamics of the object system directly and introducing an improved self-defined exciting strategy, the grey box neural network model for normal mode can be obtained to generate diagnosis residuals online as an estimator so that unmatched dynamics can be avoided. Based on the GBNNM model, faults can be detected and estimated but fault models are not essential.Finally, how to develop the software and hardware in order to validate the proposed FD mentioned above is studied. A real-time simulation hardware platform based on xPC Target, which is used to simulate the SACS, is constructed. And then with referring to some type three-axis earth-oriented satellite, a complete SACS model is designed to inject faults and simulate the fault’s behavior. Based on the software/hardware platform, faults mechanism can be analyzed and FD can be validated.更多还原