基于状态估计的多方法融合的故障预测算法研究

发布时间：2019-06-13 23:19

【摘要】：随着人们对系统性能的安全性和可靠性要求的不断提高,人们希望在系统仅仅出现微小的异常变化时,就能获得故障未来发展变化的信息,从而控制故障的持续发展,进一步减少事故发生的可能。与传统的计划维护体制相比,基于故障预测的预测维修技术能够提高设备的使用率,从而减少维护费用和生产成本。因此,研究和发展故障预测技术具有重要的理论意义和应用价值。故障预测算法通常先通过状态的监控信息和已知的历史运行数据等对系统的状态做出估计,然后对系统状态的演变趋势进行预测。基于此,本文的研究内容分为两个部分,第一部分分析了系统的非线性、模型不确定性、多重测量丢失、传感器的采样方式等因素对系统性能的影响,研究了几类随机的不确定性动态系统的状态估计问题；第二部分考虑在已知系统解析模型时,基于机理模型的方法可以有效地预估系统的运行状态,但对复杂工业系统来说依据机理建立精确的数学模型十分困难。数据驱动的方法可以利用系统的离、在线数据和其他知识,实现对故障的预报和诊断。但数据驱动的故障预报方法除了需要系统正常运行的历史数据之外还需要系统非正常运行时的数据,要获得系统故障状态下的运行数据常需要极高的代价,甚至是灾难性的。本文研究了融合两类方法的合适框架,充分使用已有的机理知识和历史数据,研究了动态数据驱动的状态估计方法来提高预测精度的问题。本文的具体研究工作概括如下：针对时变随机的非线性系统状态估计问题,综合考虑了乘性噪声、参数不确定性和多重测量数据丢失对滤波器性能的影响,设计了最小方差意义下的具有迭代形式的滤波器,使得该滤波器不仅对系统参数的不确定性具有鲁棒性,对滤波器参数的变化也有非脆弱性；针对一类基于事件驱动采样控制、带有多重测量丢失和不确定性的非线性系统,根据最小方差准则,研究了滤波器的设计问题,在每一个采样时刻,通过设计合适的滤波器增益矩阵实现最小化状态估计误差协方差的上界。所采用的滤波器具有迭代递推的形式,能够在线的进行计算。同时,对系统估计误差进行了分析,利用随机分析理论,在一定条件下,证明了估计误差的均方有界性。针对基于模型的滤波方法过度依赖所建立的系统模型,而数据驱动方法的预测稳定性和准确性较差的问题,提出了基于无迹卡尔曼滤波和相关向量机融合的预测方法。综合考虑长期和短期数据对未来趋势的影响,将相关向量机预测得到的未来时刻的滤波器残差项与短期的残差项数据进行动态加权,并代入到无迹卡尔曼滤波器之中实现了对滤波器的动态调整和预测更新,从而提高了预测的准确性。针对中／长时间跨度的预测精度较低的问题,同时,改善随着预测步长的增加数据驱动方法的预测值误差增大对融合算法的不利影响,在融合方法中通过引入自适应的权重项,从而修正了数据驱动算法的预测值对滤波更新的作用。此外,考虑到随着预测的测量值的变化,系统测量噪声在预测过程中是不断变化的,采用系统测量噪声的动态更新算法实现对系统模型的更新,从而进一步提高预测的准确性。最后,对全文的研究工作进行了总结,并对下一步的研究课题进行了展望。
[Abstract]:With the improvement of the safety and reliability requirements of the system performance, it is hoped that the information of the future development of the fault can be obtained when the system has only slight abnormal changes, so as to control the continuous development of the fault and further reduce the possibility of the accident. Compared with the traditional scheme maintenance system, the prediction and maintenance technology based on the fault prediction can improve the utilization rate of the equipment, thereby reducing the maintenance cost and the production cost. Therefore, the research and development of the fault prediction technology is of great theoretical and practical value. The fault prediction algorithm usually estimates the state of the system firstly through the monitoring information of the state and the known historical operation data, and then the evolution trend of the system state is predicted. Based on this, the research contents in this paper are divided into two parts. The first part analyzes the influence of the system's non-linearity, the model uncertainty, the multiple measurement and the sampling way of the sensors on the system performance, and studies the state estimation of several kinds of stochastic uncertain dynamic systems. In the second part, when the model of the known system is known, the operation state of the system can be estimated effectively based on the method of the mechanism model, but it is very difficult to set up an accurate mathematical model for the complex industrial system. The data-driven method can utilize the off-line, on-line data and other knowledge of the system to realize the prediction and diagnosis of the fault. However, in addition to the historical data which needs the normal operation of the system, the data-driven fault prediction method also needs the data in the abnormal operation of the system, so that the operation data under the system fault condition often needs extremely high cost, and is even catastrophic. In this paper, the proper framework of two kinds of methods is studied, the existing mechanism knowledge and historical data are fully used, and the state estimation method of dynamic data driving is studied to improve the prediction precision. The specific research work in this paper is summarized as follows: for the time-varying random nonlinear system state estimation problem, the effect of multiplicative noise, parameter uncertainty and multiple measurement data loss on the performance of the filter is comprehensively considered, and the filter with the iteration form is designed under the meaning of the minimum variance. so that the filter not only has the robustness to the uncertainty of the system parameters, but also has non-vulnerability to the change of the filter parameters, The design problem of the filter is studied. At each sampling time, the upper bound of the error covariance of the minimum state estimation error is realized by designing the appropriate filter gain matrix. The filter used has the form of iterative recursion, and can be calculated on-line. At the same time, the estimation error of the system is analyzed, and the mean-square boundedness of the estimation error is proved under certain conditions by using the random analysis theory. Aiming at the problem that the model-based filtering method is over-dependent on the established system model, and the prediction stability and the accuracy of the data driving method are poor, a method for predicting the fusion of the unscented Kalman filter and the correlation vector machine is proposed. the influence of long-term and short-term data on the future trend is comprehensively considered, the filter residual term of the future time obtained by the correlation vector machine is dynamically weighted with the short-term residual term data, and the dynamic adjustment and the prediction update to the filter are realized in the unscented Kalman filter, So as to improve the accuracy of the prediction. Aiming at the problem that the prediction accuracy of the medium/ long time span is low, meanwhile, the adverse effect of the prediction value error of the data driving method with the increase of the prediction step size on the fusion algorithm is improved, and the adaptive weight term is introduced in the fusion method, So that the effect of the prediction value of the data driving algorithm on the filtering update is corrected. In addition, taking into account the change of the measured value of the system, the measurement noise of the system is constantly changing in the prediction process, and the updating of the system model is realized by adopting a dynamic updating algorithm of the system measurement noise, so that the accuracy of the prediction is further improved. Finally, the research work of the full text is summarized, and the next research project is expected.
【学位授予单位】：华中科技大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP277

【相似文献】