A stochastic modelling development to system state prediction of high value, high risk systems subject to condition monitoring

状态监测在工业中越来越受欢迎，现在在状态监测硬件和软件上花费了大量的资金。然而，值得注意的是，尽管维护活动显著增加，状态监测设备的数量和复杂性迅速增加，但系统仍然出现故障。为什么会这样？最大的原因是维护工程师缺乏可靠的预测方法。该项目的目的是为故障检测、预测和随后的维护决策开发一种建模方法。我们采用的关键技术是隐马尔可夫模型（HMM）。它是一种广泛应用于语音识别和图像分割的技术。这里我们假设被监测的系统根据时间/年龄相关的马尔可夫过程恶化，但其状态是不可观察的。我们进一步假设观测到的监测参数受到系统的底层状态的影响，而不是相反。利用递归滤波技术建立了基于观测历史信息的初始故障检测和预测模型。该模型将在基于状态的维护决策支持中发挥重要作用，如果该模型被证明是有效的，将为英国工业节省数百万美元。

Condition monitoring is growing in popularity In industry with considerable sums now being spent on condition monitoring hardware and software. It is noted however that despite the significant rise In the profile of maintenance activities, and a burgeoning in the numbers and sophistication of condition monitoring equipment, systems continue to fail. Why is this? The single largest contributing factor Is that maintenance engineers lack a reliable way of prognosis. The aim of the project is to develop a modelling approach for fault detection, prognosis and subsequently maintenance decision making. The key technique we adopt Is what called a Hidden Markov Model (HMM) . It is a technique widely used in speech recognition and image segmentation.Here we assume the system monitored deteriorates according to a time/age dependent Markov process, but its state is unobservable. We furtherassume that the observed monitoring parameters is influenced by the underlying state of the system with random noise but not vice versus. A recursive filtering techniques is used to establish the initial fault detection and prognosis model based observed past history information. The model proposed will play a major role in condition based maintenance decision support, which in turn will save millions in UK industry if it proves to be valid.