【作者】 唐雅娟；

【导师】 张为华；

【作者基本信息】 国防科学技术大学 ， 航空宇航科学与技术， 2007， 博士

【摘要】 本文针对发动机试验领域对传感器数据证实技术的迫切需求,研究传感器数据证实的方法与应用系统实现的关键技术。针对试验数据中脉冲型噪声去除问题,设计了中心加权等幂中值滤波器及其快速实现算法。设计了模拟发动机故障、性能退化和瞬变工况并包含野值与随机白噪声的传感器信号,测试了中值滤波器保持信号真实突变特征并去除脉冲噪声的性能及计算效率,将其应用于涡轮试验数据预处理取得良好效果。针对线性关联多通道信号偏差检测问题,发展了主元—残差子空间方法。导出了偏差幅值与子空间参数对检测统计量影响的定量关系,发展了利用异常数据在多尺度分解下的统计量相对变化特征检测小幅值偏差的多尺度主元—残差子空间方法。应用于涡轮试验数据分析,表明该方法可减小偏差检测滞后性并降低偏差检测虚警率。针对具有非线性和时变性关系的涡轮试验多通道信号偏差检测问题,建立了自关联神经网络估计器和预报器方法。分析了自关联神经网络结构及输入—输出参数与数据内在物理关联性之间的关系,研究提出了不同输入参数的涡轮试验传感器数据估计器和预报器,可实现偏差检测、分离及数据重构。针对多源证据融合问题,建立了贝叶斯信度网络在发动机及其部件试验传感器数据证实中的应用方法,涉及传感器状态和检验关系式不确定性信息表达方法,根据传感器与检验关系式集合自动建立贝叶斯信度网络、计算可信度概率及更新网络的算法。分析了在线证实的实时性问题,给出了贝叶斯信度网络建立原则。针对高压涡轮试验数据证实系统开发需要,建立了流体和机械运动参数之间的稳态关联方程,分析了基于物理模型的解析冗余方法可用性,给出了时域滑动数据窗相关方法在涡轮试验数据证实中的应用方案。解决了传感器数据证实系统涉及的检验关系式数目、残差阈值、单周期决策逻辑、多周期决策策略、系统可放大性等问题;给出了传感器选择、关系式建立途径及系统调试方法,设计实现了传感器数据证实网络自动生成系统和实时运行内核。建立了组合解析冗余、统计相关检验与贝叶斯信度网络融合证据的传感器数据证实方案,实现了高压涡轮试验传感器数据证实系统。该系统具有事后证实功能,测试表明系统具有实用价值。更多还原

【Abstract】 Sensor data validation is severely concerned by engineers of engine test domain, which calls forth the study on methods and system realization of sensor data validation by soft-computing technique for engine tests in this paper.A center weighted idempotent median (CWIM) filter was introduced to removing high-amplitude impulsive noise in signals of gas turbine test, and a fast algorithm for the specific CWIM filter was devised. The filtering function and algorithmic efficiency were evaluated using a test signal, which representing the sensor response to engine abrupt faults, linear deterioration and/or regime transients, while contaminated with random noise and high-amplitude impulsive noise. Results confirmed that median filter is good at removing impulsive noise while preserving real steep edges in signals, also expected effect was obtained when applied to gas turbine test data.A subspace model based approach was presented for detection of errors in signals with linear correlation. The normal subspace model was generated through principal component analysis (PCA), and statistics in conventional principal component subspace and residual subspace were quantitatively connected with error magnitude and subspace characteristics. The characteristics of statistics varying with data faults development in multi-scale PCA was analyzed, and a multi-scale subspace based method was developed for detection of small errors. The time lag of error detection was shortened, and false alarm rate was reduced when applied to practical gas turbine test data.A neural network model based approach was presented to handle the bias detection and correction problem with data correlated by nonlinear and time-varying relations in gas turbine test. A special architecture of the auto-associative neural network was defined with different input and output parameters. Novel estimators and predictors based on auto-associative neural network were devised and evaluated with practical test data. A scheme integrating operations of error detection, fault isolation and data reconstruction was presented based on auto-associative neural network.A method for fusing evidence information using Bayesian belief network was introduced into sensor data validation. Uncertainty expression of sensor state and relations in engine and its components test were defined. The algorithms for automatic generation of the Bayesian belief network files, belief probability calculation and network update after a abruption of one faulty sensor were developed. Real-time running feasibility was analyzed, and criteria for Bayesian belief network evaluation was presented.The mathematical models for fluid and mechanical characteristics of high pressure gas turbine test system in steady regimes were developed. The effectiveness of analytical redundancy technique based on first principles models was evaluated. A scheme based on moving window technique that continuously computes auto-correlation function of samples of sensor data in time domain was devised. A solution for validating gas turbine test sensor data was presented, which answered critical problems about the construction of a data validation system, including check relation number, residual thresholds, single cycle decision logic, multi-cycle decision strategy, system scaleable capacity to any sensor set, etc. The methods for sensor selection, relations definition and system test were presented in detail. A sensor validation network development system and a real-time kernel was developed in software. A prototype system was realized, which well demonstrated the versatility and effectiveness for post validation of practical high pressure gas turbine test data.更多还原