Study On Multi-Action Dynamic Maintenance Policy Based On Non-Homogeneous Markov Process

With rapid development of computer and electronic technology, the development of production equipment trends to the direction of high speed, high load and high automation. But the serious incidents caused by equipment failures have increased significantly, and have a serious impact on society. As equipments run, system state, operation cost, and reliability of maintenance actions will vary. So the traditional periodic policy will result in maintenance shortage or overage, and dynamic prevent policy based on condition has been a popular direction of research.The object of this research is a whole machine equipped with inspection equipments. The basic flow is:Firstly, introduce the concept of whole health index, and establish a Whole Health Index model. Through standardizing the real-time data, we can get the whole health index of this machine, and then classify the operation situation into several discrete states;Secondly, because of the self-study and function fitting ability,neural network is introduced to estimate the distribution function, which the degrees of state deterioration ( differences between the states before and after transitions) follow, from historical state database. Then state transition probability matrices varying throughout the running can be obtained.And then, with the state transition probability matrices, this research adopts a discrete time non-homogeneous Markovian multi-state deteriorating model to describe the state transitions of a periodically inspected and maintained machine. The model expresses that: machine states are deteriorating (via upper triangular transition probability matrices), and machines are aging (via non-homogeneous transition probabilities). By further assuming that multiple actions are available, each action with its own risk (a maintenance may not achieve its intended result), and the actions time cann't be neglected. A downtime loss factor is employed to transfer down time into loss cost, and the dynamic programming algorithm is developed to determine the optimal action according to the present value of minimal expected total cost during a time interval in the near future.The result of numerical analysis showed that the cost can be reduced remarkably in the dynamic policy under various combination of several parameters compared with traditional periodic maintenance policy.