面向故障预测的多源不确定性建模方法研究

This research mainly focuses on a data-driven prognostics and health management (PHM) framework for remaining useful life (RUL) prediction and prognostics of spacecrafts, in which the prognostics uncertainty from multiple sources is emphasized. The theory and algorithms for prognostics uncertainty identification, quantification, evaluation and fusion will be studied in depth. Firstly, the uncertainty quantification metrics and methods of prognostics algorithms will be studied, as a result, prognostics algorithms with uncertainty can be evaluated scientifically. The uncertainty analysis of the samples and the model parameters will be conducted and modeled to optimize the prognostics algorithms. Here, we will mainly focus on the uncertainty modeling of dynamic failure threshold and multiple degradation modes. Secondly, uncertainty fusion for integrating hybrid data-driven prognostics algorithms will be concentrated, which contains two aspects: homogeneous fusion and heterogeneous fusion. The uncertainty improvement will be implemented for those prognostics algorithms providing only point estimates of RUL. On the other hand, the mixture of distributions and Bayesian networks will be applied to fuse prognostics algorithm. Consequently, the desired uncertainty fusion as well as the optimization for various prognostics algorithms can be achieved. Moreover, the uncertainty minimization via prognostics algorithms integration will be realized. The uncertainty in inputs, data, models, etc. can be controlled flexibly in the hybrid data-driven prognostics framework. Finally, we will evaluate the uncertainty theory and algorithms with simulated data sets, prognostics open resource data sets, and actual testing and in-orbit data sets of components and subsystems of satellites. This work will result in the modeling representation of multi-sources uncertainties and minimization of uncertainty fusion for RUL estimation. In addition, the restriction with existing prognostics algorithms that can not accommodate multi-sources uncertainties can be solved. Furthermore, it will bring about significant improvement in space applications in both theoretical and technological aspects.

针对航天器故障预测和寿命预计的不确定性问题，研究故障预测多源不确定性建模的理论和算法，涵盖不确定性识别、量化、评估和融合。第一，研究故障预测不确定性因素建模，建立不确定性评价的参数指标体系，实现样本和模型参数的不确定性影响分析和建模优化，重点突破动态失效阈值和多退化模式条件下的不确定性建模难题；第二，研究故障预测算法集成过程中不确定性融合，针对同构和异构预测算法的不确定性融合两大类基本问题，研究预测算法的不确定特性改进、基于混合概率分布和贝叶斯网络的不确定性融合方法，实现不确定性融合的优化及最小化；最后，基于仿真数据集、故障预测公开数据集以及实际卫星部件和子系统的地面试验和在轨状态监测数据，开展不确定性建模方法体系的应用验证和综合评估。研究将实现故障预测不确定性的模型化表征及不确定性融合的优化，突破现有方法难以支持多种不确定性因素的局限，为未来航天器的实际应用奠定不确定性理论和算法基础。

针对航天/航空器关键部件、子系统故障预测和寿命预计的不确定性问题，研究故障预测多源不确定性建模的理论和算法，涵盖不确定性识别、量化、评估和融合。首先，针对缺少故障预测的多源不确定性建模的理论和方法框架的问题，开展了故障预测的不确定性建模的理论和算法研究，构建了面向航天/航空器关键部件故障预测的不确定性建模研究体系；第二，针对故障预测建模中健康状态表征易受工况影响存在适应性差的问题，开展了样本、模型参数变化条件下健康因子构建模型研究，结合航空器关键部件实际需求，提出基于贝叶斯更新的健康因子构建方法，实现不确定性条件下健康状态的准确表征。第三，针对固定失效阈值条件下性能预测方法难以准确评估部件或设备性能的问题，开展动态失效阈值条件下的不确定性建模方法研究，实现采样数据的波动范围建模和不同分布序列的预测区间优化。第四，针对传统寿命预测建模中未考虑多种退化模式条件与RUL预测的结合问题，开展多退化模式的不确定性建模方法研究，提出基于贝叶斯模型平均的概率集成RUL预测方法，实现预测精度及稳定性的大幅提升。第五，针对不同退化特性与数据特点引起的故障预测方法性能和适应能力差异较大的问题，开展多预测模型融合的不确定性建模方法研究，实现多个同质及异质预测模型的融合以及集成模型的不确定性融合和优化。最后，在数据仿真验证和公开数据集验证的基础上，实现面向航天/航空器关键部件、子系统不确定性模型的应用验证和评估。.依托本课题研究，项目组实现了故障预测不确定性的模型化表征及不确定性融合的优化，突破了现有方法难以支持多种不确定性因素的局限，为航天/航空器寿命预测和健康管理、大规模监测数据分析奠定不确定性理论和算法基础，并为大型卫星星座与无人机集群的复杂数据建模与分析提供强有力的技术支撑，推动航空航天领域数据价值向多角度应用的快速转化，为我国由航天大国向航天强国过度提供坚实的技术储备。