【作者】 夏鲁瑞；

【导师】 胡茑庆；

【作者基本信息】 国防科学技术大学 ， 机械工程， 2010， 博士

【摘要】 从人类发射第一颗人造地球卫星以来,航天技术的发展十分迅速并取得了举世瞩目的成就,但同时也对液体火箭发动机提出了更高的要求,迫切需要提高其可靠性和安全性。涡轮泵是液体火箭发动机的重要组件,由于工作在极端恶劣的物理环境中,一旦出现故障,会迅速扩展演化,严重威胁发动机的安全。涡轮泵健康监控技术及系统的目标是及时发现故障,并结合相应的控制措施,以有效缓解故障劣化,减小故障对发动机乃至整个地面试车台的影响。因此,涡轮泵健康监控关键技术及系统研究对于液体火箭发动机地面综合试车的安全保障具有重要意义。本文针对目前液体火箭发动机地面试车中涡轮泵健康监控存在的主要问题,在国家863课题和国家自然科学基金项目的支持下,采用理论与应用相结合的方法,开展涡轮泵健康监控关键技术及系统研究。论文完成的主要工作包括：1.详细分析了液体火箭发动机健康监控技术及系统的研究现状,论述了其发展方向,探讨了目前我国液体火箭发动机涡轮泵健康监控需要突破的一些主要关键技术,即启动过程监控技术、稳态工作过程突变故障检测技术、试车后海量数据信息提取技术以及健康监控系统构建技术。2.研究并提出涡轮泵启动过程信号特征提取与监控方法(1)针对涡轮泵启动过程转速变化造成的振动特征提取困难,研究并实现了基于数字重采样的计算阶比分析方法,有效计算和提取了转速和阶比特征。(2)提出了基于转速特征和阶跟踪特征的涡轮泵启动过程信号监控方法,并通过涡轮泵历史试车数据验证了方法的有效性。(3)结合支持向量回归技术,提出了基于非线性模型和转速安全带的涡轮泵启动过程故障检测方法,并通过涡轮泵试车数据对方法进行了验证。发动机热试车数据验证表明,基于数字重采样的计算阶比分析方法以及基于转速和阶比特征的监控方法可在启动过程有效识别出涡轮泵存在的早期故障,为涡轮泵工作全过程的故障检测提供了一条有效途径。3.研究并提出基于多变换域特征的涡轮泵稳态过程突变故障自适应检测算法(1)在分析涡轮泵突变故障机理与振动特征的基础上,选择了可有效反映突变故障的多个时域和频域特征,提出了转速波动情况下特征频率的阶域提取方法。(2)研究了特征阈值模型和自适应计算理论,根据工程实际,改进与实现了基于多时域特征的突变故障自适应检测算法。(3)研究了频域红线算法的原理,实现了特征频率成分的幅值跟踪,提出了基于特征频率的突变故障自适应红线检测算法。发动机热试车数据验证表明,阶域提取方法可有效提取稳定的特征频率成分,为解决转速波动造成的特征谱线提取困难提供了一条有效的途径；基于多时域特征的自适应检测算法能避免振动过小引起的误报警,与传统方法相比,更适用于涡轮泵稳态过程的突变故障检测；而基于特征频率的自适应红线检测算法可以有效检测出涡轮泵叶片脱落等突变故障。4.研究并提出基于流形学习的涡轮泵试车后健康分析方法(1)在深入研究流形学习的数学描述、基本思想和主要方法的基础上,提出了一种从海量试车数据中提取有用信息的新策略。(2)描述了数据的流形特征,分析了其用于涡轮泵试车数据异常识别的有效性,提出了一种基于流形特征的海量数据异常识别算法,并通过仿真和涡轮泵试车数据对算法进行了验证。研究表明,通过本文算法提取海量数据的流形特征,数据量可大大约简,同时利用特征空间中流形不同区域的差别,可实现涡轮泵正常和异常状态的识别。5.研究了涡轮泵健康监控系统构建技术(1)针对监控系统的小型化需求,提出了基于数字信号处理器(DSP)的新型嵌入式监控系统解决方案,并进行了软硬件设计,构建了实时健康监控子系统。(2)集成了多变换域分析和维数约简方法,构建了涡轮泵试车后健康分析子系统。研究表明,基于DSP的原型子系统具备涡轮泵健康状态的实时监控能力,能够实时计算转速、提取振动信号的时域和频域特征,并进行故障检测；试车后健康分析子系统能更深入地分析和进一步确认涡轮泵的运行状态。综上所述,本文对液体火箭发动机涡轮泵健康监控关键技术及系统进行了深入的研究工作：针对启动过程信号特征提取的困难,提出了涡轮泵启动过程信号监控方法；在提取稳态过程振动信号多变换域特征的基础上,提出了涡轮泵稳态过程突变故障的自适应检测算法；深入研究流形学习方法,提出了基于流形特征的涡轮泵海量试车数据异常识别算法；设计和构建了涡轮泵健康监控原型系统,并通过了历史试车数据的验证和现场热试车的考核。本文的研究工作,可为开展液体火箭发动机和可重复使用发动机健康监控技术研究及系统研制,提供学术参考和工程借鉴。更多还原

【Abstract】 Since the first man-made earth satellite revolves round the earth, the space technology is developing rapidly and its achievements have attracted worldwide attention. Meanwhile, it is necessary to enhance the reliability and security of liquid-propellant rocket engine (LRE). Turbopump is the important component of LRE. Because of its execrable working circumstance, its faults evolve very fast and severely endanger the safety of LRE. The goal of health monitoring technology and system for turbopump is to detect faults in time. With various control measures, the system can mitigate faults deterioration and diminish their impact on engine and ground test-bed. Thus, the research on health monitoring technology and system for turbopump is of importance to guarantee the safety of LRE in the synthetic ground test.Aiming at the existing principal problem of turbopump health monitoring in LRE ground test, granted by National 863 Project and Natural Science Foundation of China (NSFC), this dissertation carries out the research on health monitoring technology and system for turbopump combining theory with application. The main contents and innovative work can be summarized as follows:1. The art of state of LRE health monitoring technology and system is analyzed in detail. Based on this, their developing trend is described and the key techniques that need to be resolved presently for LRE turbopump health monitoring are investigated. These techniques consist of the monitoring technology in start-up process, the abrupt fault detection technology in steady working process, the information extraction technology for post-test mass data and the system construction technology for turbopump health monitoring.2. The feature extraction and monitoring methods for turbopump start-up process signal are studied and proposed.(1) The variation of rotational speed in turbopump start-up process presents difficulty to vibration feature extraction. For this reason, the computed order analysis method based on digital resample is studied and realized, which can compute the speed and extract order features effectively.(2) The turbopump start-up process monitoring methods based on the speed and order features are brought forward. With turbopump historical test data, the validity of the methods is verified.(3) Associating with support vector regression, the turbopump start-up process fault detection methods based on nonlinear model and speed safety belt are presented and verified with turbopump test data.Validated by LRE firing test data, the computed order analysis method via digital resample and the monitoring methods based on the speed and order features can recognize availably the initial failure in turbopump start-up process, which supplies an effective approach to the fault detection in the entire working process of turbopump.3. The adaptive detection algorithms based on multiple domain features are studied and proposed for the turbopump abrupt faults in steady working process.(1) After the mechanism and vibration characteristics of turbopump abrupt faults are analyzed, the multiple effective features in time domain and frequency domain are selected. Furthermore, the order domain extraction method of the feature frequency in speed fluctuating condition is put forward.(2) The feature threshold model and adaptive computation theory are studied. According to actual requirement, the adaptive detection algorithm based on multiple time domain features for abrupt faults is improved and realized.(3) The red-line algorithm in frequency domain is studied and the amplitude tracking of feature frequency components is realized. Then, the adaptive red-line detection algorithm based on feature frequency for abrupt faults is proposed.Validated by LRE firing test data, the order domain extraction method can effectively extract the stable feature frequency components, which supplies an effective approach to solve the difficulty of feature extraction as a result of speed fluctuation. The adaptive detection algorithm based on multiple time domain features can avoid the false alarm proceeding from undersized vibration. Compared with traditional methods, it is more applicable to the abrupt fault detection in turbopump steady process. Meanwhile, the adaptive red-line detection algorithm based on feature frequency can detect effectively the abrupt fault of turbopump blade abscission.4. The turbopump post-test health analysis method based on manifold learning is studied and proposed.(1) After the mathematical description, fundamental conception and primary methods of manifold learning are studied in depth, a new strategy for the useful information extracting from mass test data is presented.(2) The manifold features of data are described and the validity of the anomaly recognition for turbopump test data is analyzed. Then, a mass data anomaly recognition algorithm based on manifold features is proposed and verified with simulation and turbopump test data.The research shows that the quantity of data can be greatly reduced after the manifold features are extracted from mass data by the presented algorithm. In the meanwhile, the recognition for turbopump normal and abnormal state can be carried out by means of the difference among manifold regions in feature space.5. The system construction technology for turbopump health monitoring is studied.(1) Aiming at the small-sized requirement for monitoring system, a new solution for the embedded monitoring system based on digital signal processor (DSP) is presented and the soft hardware design is carried out. Then, the real-time health monitoring subsystem is built up.(2) The turbopump post-test health analysis subsystem is constructed by integrating multiple domain analysis with dimensionality reduction method.The research shows that the subsystem based on DSP has the real-time capability of monitoring turbopump health condition. It is able to compute promptly the speed, extract the vibration features in time domain and frequency domain, and then detect faults. The post-test health analysis subsystem can further analyze and confirm the running state of turbopump.In a summary, the key technology and system for LRE turbopump health monitoring are studied thoroughly. Aiming at the difficulty of feature extraction in start-up process, the monitoring methods for turbopump start-up process signal are proposed. The multiple domain features of vibration signal are extracted, and then the adaptive detection algorithms are presented for the turbopump abrupt faults in steady working process. After the manifold learning methods are studied in depth, the anomaly recognition algorithm based on manifold features for turbopump mass test data is brought forward. Finally, the turbopump health monitoring system is designed and built, which is verified with historical test data and LRE firing test data. It is believed that these conclusions can provide academic reference and engineering example for research and development of LRE and reusable engine health monitoring technology and system.更多还原