Monotone optimal preventive maintenance policies for stochastically failing equipment

This paper examines various models for maintenance of a machine operating subject to stochastic deterioration. Three alternative models are presented for the deterioration process. For each model, in addition to the replacement decision, the option exists of performing preventive maintenance. The effect of this maintenance is to “slow” the deterioration process. With an appropriate reward structure imposed on the processes, the models are formulated as continuous time Markov decision processes. the optimality criterion being the maximization of expected discounted reward earned over an infinite time horizon. For each model conditions are presented under which the optimal maintenance policy exhibits the following monotonic structure. First, there exists a control limit rule for replacement. That is, there exists a number i* such that if the state of machine deterioration exceeds i* the optimal policy replaces the machine by a new machine. Secondly, prior to replacement the optimal level of preventive maintenance is a nonincreasing function of the state of machine deterioration. The conditions which guarantee this result have a cost/benefit interpretation.     

Optimal joint preventive maintenance and production policies

We study joint preventive maintenance (PM) and production policies for an unreliable production‐inventory system in which maintenance/repair times are non‐negligible and stochastic. A joint policy decides (a) whether or not to perform PM and (b) if PM is not performed, then how much to produce. We consider a discrete‐time system, formulating the problem as a Markov decision process (MDP) model. The focus of the work is on the structural properties of optimal joint policies, given the system state comprised of the system's age and the inventory level. Although our analysis indicates that the structure of optimal joint policies is very complex in general, we are able to characterize several properties regarding PM and production, including optimal production/maintenance actions under backlogging and high inventory levels, and conditions under which the PM portion of the joint policy has a control‐limit structure. In further special cases, such as when PM set‐up costs are negligible compared to PM times, we are able to establish some additional structural properties. © 2005 Wiley Periodicals, Inc. Naval Research Logistics, 2005.     

Simultaneous determination of production and maintenance schedules using in‐line equipment condition and yield information

In many manufacturing environments, equipment condition has a significant impact on product quality, or yield. This paper presents a semi‐Markov decision process model of a single‐stage production system with multiple products and multiple maintenance actions. The model simultaneously determines maintenance and production schedules, accounting for the fact that equipment condition affects the yield of each product differently. It extends earlier work by allowing the expected time between decision epochs to vary by both action and machine state, by allowing multiple maintenance actions, and by treating the outcome of maintenance as less than certain. Sufficient conditions are developed that ensure the monotonicity of both the optimal production and maintenance actions. While the maintenance conditions closely resemble previously studied conditions for this type of problem, the production conditions represent a significant departure from earlier results. The simultaneous solution method is compared to an approach commonly used in industry, where the maintenance and production problems are treated independently. Solving more than one thousand test problems confirms that the combination of both features of the model—accounting for product differences and solving the problems simultaneously—has a significant impact on performance. © 2007 Wiley Periodicals, Inc. Naval Research Logistics, 2008     

Minimizing the false alarm rate in systems with transient abnormality

We consider a stochastic partially observable system that can switch between a normal state and a transient abnormal state before entering a persistent abnormal state. Only the persistent abnormal state requires alarms. The transient and persistent abnormal states may be similar in appearance, which can result in excess false alarms. We propose a partially observable Markov decision process model to minimize the false alarm rate, subject to a given upper bound on the expected alarm delay time. The cost parameter is treated as the Lagrange multiplier, which can be estimated from the bound of the alarm delay. We show that the optimal policy has a control‐limit structure on the probability of persistent abnormality, and derive closed‐form bounds for the control limit and present an algorithm to specify the Lagrange multiplier. We also study a specialized model where the transient and persistent abnormal states have the same observation distribution, in which case an intuitive “watchful‐waiting” policy is optimal. © 2016 Wiley Periodicals, Inc. Naval Research Logistics 63: 320–334, 2016     

Optimizing usage and maintenance decisions for k‐out‐of‐n systems of moving assets

We consider an integrated usage and maintenance optimization problem for a k‐out‐of‐n system pertaining to a moving asset. The k‐out‐of‐n systems are commonly utilized in practice to increase availability, where n denotes the total number of parallel and identical units and k the number of units required to be active for a functional system. Moving assets such as aircraft, ships, and submarines are subject to different operating modes. Operating modes can dictate not only the number of system units that are needed to be active, but also where the moving asset physically is, and under which environmental conditions it operates. We use the intrinsic age concept to model the degradation process. The intrinsic age is analogous to an intrinsic clock which ticks on a different pace in different operating modes. In our problem setting, the number of active units, degradation rates of active and standby units, maintenance costs, and type of economic dependencies are functions of operating modes. In each operating mode, the decision maker should decide on the set of units to activate (usage decision) and the set of units to maintain (maintenance decision). Since the degradation rate differs for active and standby units, the units to be maintained depend on the units that have been activated, and vice versa. In order to minimize maintenance costs, usage and maintenance decisions should be jointly optimized. We formulate this problem as a Markov decision process and provide some structural properties of the optimal policy. Moreover, we assess the performance of usage policies that are commonly implemented for maritime systems. We show that the cost increase resulting from these policies is up to 27% for realistic settings. Our numerical experiments demonstrate the cases in which joint usage and maintenance optimization is more valuable. © 2017 Wiley Periodicals, Inc. Naval Research Logistics 64: 418–434, 2017     

Structured maintenance policies on interior sample paths

We examine the problem of adaptively scheduling perfect observations and preventive replacements for a multi‐state, Markovian deterioration system with silent failures such that total expected discounted cost is minimized. We model this problem as a partially observed Markov decision process and show that the structural properties of the optimal policy hold for certain non‐extreme sample paths. © 2007 Wiley Periodicals, Inc. Naval Research Logistics, 2007     

Technical note: A computationally efficient algorithm for undiscounted Markov decision processes with restricted observations

We present a computationally efficient procedure to determine control policies for an infinite horizon Markov Decision process with restricted observations. The optimal policy for the system with restricted observations is a function of the observation process and not the unobservable states of the system. Thus, the policy is stationary with respect to the partitioned state space. The algorithm we propose addresses the undiscounted average cost case. The algorithm combines a local search with a modified version of Howard's (Dynamic programming and Markov processes, MIT Press, Cambridge, MA, 1960) policy iteration method. We demonstrate empirically that the algorithm finds the optimal deterministic policy for over 96% of the problem instances generated. For large scale problem instances, we demonstrate that the average cost associated with the local optimal policy is lower than the average cost associated with an integer rounded policy produced by the algorithm of Serin and Kulkarni Math Methods Oper Res 61 (2005) 311–328. © 2008 Wiley Periodicals, Inc. Naval Research Logistics 2009     

Optimal maintenance policies for a safety‐critical system and its deteriorating sensor

We consider the integrated problem of optimally maintaining an imperfect, deteriorating sensor and the safety‐critical system it monitors. The sensor's costless observations of the binary state of the system become less informative over time. A costly full inspection may be conducted to perfectly discern the state of the system, after which the system is replaced if it is in the out‐of‐control state. In addition, a full inspection provides the opportunity to replace the sensor. We formulate the problem of adaptively scheduling full inspections and sensor replacements using a partially observable Markov decision process (POMDP) model. The objective is to minimize the total expected discounted costs associated with system operation, full inspection, system replacement, and sensor replacement. We show that the optimal policy has a threshold structure and demonstrate the value of coordinating system and sensor maintenance via numerical examples. © 2017 Wiley Periodicals, Inc. Naval Research Logistics 64: 399–417, 2017     

基于隐马尔可夫模型的IDS程序行为异常检测

提出一种新的基于隐马尔可夫模型的程序行为异常检测方法,此方法利用系统调用序列,并基于隐马尔可夫模型来描述程序行为,根据程序行为模式的出现频率对其进行分类,并将行为模式类型同隐马尔可夫模型的状态联系在一起。由于各状态对应的观测值集合互不相交,模型训练中采用了运算量较小的序列匹配方法,与传统的Baum Welch算法相比,训练时间有较大幅度的降低。考虑到模型中状态的特殊含义以及程序行为的特点,将加窗平滑后的状态序列出现概率作为判决依据。实验表明,此方法具有很高的检测准确性,其检测效率也优于同类方法。     

On‐line parameter estimation for a partially observable system subject to random failure

In this paper, we study the on‐line parameter estimation problem for a partially observable system subject to deterioration and random failure. The state of the system evolves according to a continuous time homogeneous Markov process with a finite state space. The system state is not observable, except for the failure state. The information related to the system state is available at discrete times through inspections. A recursive maximum likelihood (RML) algorithm is proposed for the on‐line parameter estimation of the model. The RML algorithm proposed in the paper is considerably faster and easier to apply than other RML algorithms in the literature, because it does not require projection into the constraint domain and calculation of the gradient on the surface of the constraint manifolds. The algorithm is illustrated by an example using real vibration data. © 2006 Wiley Periodicals, Inc. Naval Research Logistics, 2006     

Model parameter estimation and residual life prediction for a partially observable failing system

We consider a partially observable degrading system subject to condition monitoring and random failure. The system's condition is categorized into one of three states: a healthy state, a warning state, and a failure state. Only the failure state is observable. While the system is operational, vector data that is stochastically related to the system state is obtained through condition monitoring at regular sampling epochs. The state process evolution follows a hidden semi‐Markov model (HSMM) and Erlang distribution is used for modeling the system's sojourn time in each of its operational states. The Expectation‐maximization (EM) algorithm is applied to estimate the state and observation parameters of the HSMM. Explicit formulas for several important quantities for the system residual life estimation such as the conditional reliability function and the mean residual life are derived in terms of the posterior probability that the system is in the warning state. Numerical examples are presented to demonstrate the applicability of the estimation procedure and failure prediction method. A comparison results with hidden Markov modeling are provided to illustrate the effectiveness of the proposed model. © 2015 Wiley Periodicals, Inc. Naval Research Logistics 62: 190–205, 2015     

State information lag markov decision process with control limit rule

In the framework of a discrete Markov decision process with state information lag, this article suggests a way for selecting an optimal policy using the control limit rule. The properties sufficient for an optimal decision rule to be contained in the class of control limit rules are also studied. The degradation in expected reward from that of the perfect information process provides a measure of the potential value of improving the information system.     

Optimal replacement of continuously degrading systems in partially observed environments

Motivated by wind energy applications, we consider the problem of optimally replacing a stochastically degrading component that resides and operates in a partially observable environment. The component's rate of degradation is modulated by the stochastic environment process, and the component fails when it is accumulated degradation first reaches a fixed threshold. Assuming periodic inspection of the component, the objective is to minimize the long‐run average cost per unit time of performing preventive and reactive replacements for two distinct cases. The first case examines instantaneous replacements and fixed costs, while the second considers time‐consuming replacements and revenue losses accrued during periods of unavailability. Formulated and solved are mixed state space, partially observable Markov decision process models, both of which reveal the optimality of environment‐dependent threshold policies with respect to the component's cumulative degradation level. Additionally, it is shown that for each degradation value, a threshold policy with respect to the environment belief state is optimal if the environment alternates between two states. The threshold policies are illustrated by way of numerical examples using both synthetic and real wind turbine data. © 2015 Wiley Periodicals, Inc. Naval Research Logistics 62: 395–415, 2015     

Optimal control of a production‐inventory system with both backorders and lost sales

We consider the optimal control of a production inventory‐system with a single product and two customer classes where items are produced one unit at a time. Upon arrival, customer orders can be fulfilled from existing inventory, if there is any, backordered, or rejected. The two classes are differentiated by their backorder and lost sales costs. At each decision epoch, we must determine whether or not to produce an item and if so, whether to use this item to increase inventory or to reduce backlog. At each decision epoch, we must also determine whether or not to satisfy demand from a particular class (should one arise), backorder it, or reject it. In doing so, we must balance inventory holding costs against the costs of backordering and lost sales. We formulate the problem as a Markov decision process and use it to characterize the structure of the optimal policy. We show that the optimal policy can be described by three state‐dependent thresholds: a production base‐stock level and two order‐admission levels, one for each class. The production base‐stock level determines when production takes place and how to allocate items that are produced. This base‐stock level also determines when orders from the class with the lower shortage costs (Class 2) are backordered and not fulfilled from inventory. The order‐admission levels determine when orders should be rejected. We show that the threshold levels are monotonic (either nonincreasing or nondecreasing) in the backorder level of Class 2. We also characterize analytically the sensitivity of these thresholds to the various cost parameters. Using numerical results, we compare the performance of the optimal policy against several heuristics and show that those that do not allow for the possibility of both backordering and rejecting orders can perform poorly.© 2010 Wiley Periodicals, Inc. Naval Research Logistics 2010     

Continuous-time markov decision processes with nonzero terminal reward

In this article we consider a continuous-time Markov decision process with a denumerable state space and nonzero terminal rewards. We first establish the necessary and sufficient optimality condition without any restriction on the cost functions. The necessary condition is derived through the Pontryagin maximum principle and the sufficient condition, by the inherent structure of the problem. We introduce a dynamic programming approximation algorithm for the finite-horizon problem. As the time between discrete points decreases, the optimal policy of the discretized problem converges to that of the continuous-time problem in the sense of weak convergence. For the infinite-horizon problem, a successive approximation method is introduced as an alternative to a policy iteration method.     

Maintaining systems with heterogeneous spare parts

We consider the problem of optimally maintaining a stochastically degrading, single‐unit system using heterogeneous spares of varying quality. The system's failures are unannounced; therefore, it is inspected periodically to determine its status (functioning or failed). The system continues in operation until it is either preventively or correctively maintained. The available maintenance options include perfect repair, which restores the system to an as‐good‐as‐new condition, and replacement with a randomly selected unit from the supply of heterogeneous spares. The objective is to minimize the total expected discounted maintenance costs over an infinite time horizon. We formulate the problem using a mixed observability Markov decision process (MOMDP) model in which the system's age is observable but its quality must be inferred. We show, under suitable conditions, the monotonicity of the optimal value function in the belief about the system quality and establish conditions under which finite preventive maintenance thresholds exist. A detailed computational study reveals that the optimal policy encourages exploration when the system's quality is uncertain; the policy is more exploitive when the quality is highly certain. The study also demonstrates that substantial cost savings are achieved by utilizing our MOMDP‐based method as compared to more naïve methods of accounting for heterogeneous spares.     

基于马尔科夫决策的目标选择策略

目标选择是军事计划的关键要素之一。基于马尔科夫决策方法,解决具有复杂目标间关联的多阶段目标选择问题。使用与或树描述目标体系各层状态间的影响关联,并以目标体系整体失效为求解目的,建立了基于离散时间MDP的多阶段打击目标选择模型。在LRTDP算法基础上提出一种启发式方法,通过判断从当前目标体系状态到达体系失效状态的演化过程中的可能资源消耗和失败概率,来提供对当前状态的评估值,该方法能有效排除问题搜索空间中不能到达体系失效目的的中间状态,压缩了由于目标间复杂关联而增长的巨大状态空间。用实验验证了该方法有效性,实验结果表明,该方法直观实用,对目标间具有复杂关联关系的目标打击决策有一定参考价值。     

Optimal production ramp-up in the smartphone manufacturing industry

Motivated by challenges in the smartphone manufacturing industry, we develop a dynamic production ramp-up model that can be applied to economically satisfy nonstationary demand for short-life-cycle products by high-tech companies. Due to shorter life cycles and more rapid evolution of smartphones, production ramp-up has been increasingly critical to the success of a new smartphone. In the production ramp-up, the key challenge is to match the increasing capacity to nonstationary demand. The high-tech smartphone manufacturers are urged to jointly consider the effect of increasing capacity and decreasing demand. We study the production planning problem using a high-dimensional Markov decision process (MDP) model to characterize the production ramp-up. To address the curse of dimensionality, we refine Monte Carlo tree search (MCTS) algorithm and theoretically analyze its convergence and computational complexity. In a real case study, we find that the MDP model achieves revenue improvement by stopping producing the existing product earlier than the benchmark policy. In synthetic instances, we validate that the proposed MCTS algorithm saves computation time without loss of solution quality compared with traditional value iteration algorithm. As part of the Lenovo production solution, our MDP model enables high-tech smartphone manufacturers to better plan the production ramp-up.     

Production to order and off‐line inspection when the production process is partially observable

This study combines inspection and lot‐sizing decisions. The issue is whether to INSPECT another unit or PRODUCE a new lot. A unit produced is either conforming or defective. Demand need to be satisfied in full, by conforming units only. The production process may switch from a “good” state to a “bad” state, at constant rate. The proportion of conforming units in the good state is higher than in the bad state. The true state is unobservable and can only be inferred from the quality of units inspected. We thus update, after each inspection, the probability that the unit, next candidate for inspection, was produced while the production process was in the good state. That “good‐state‐probability” is the basis for our decision to INSPECT or PRODUCE. We prove that the optimal policy has a simple form: INSPECT only if the good‐state‐probability exceeds a control limit. We provide a methodology to calculate the optimal lot size and the expected costs associated with INSPECT and PRODUCE. Surprisingly, we find that the control limit, as a function of the demand (and other problem parameters) is not necessarily monotone. Also, counter to intuition, it is possible that the optimal action is PRODUCE, after revealing a conforming unit. © 2007 Wiley Periodicals, Inc. Naval Research Logistics, 2007     

Optimal policies for stochastic clearing systems with time-dependent delay penalties

We study stochastic clearing systems with a discrete-time Markovian input process, and an output mechanism that intermittently and instantaneously clears the system partially or completely. The decision to clear the system depends on both quantities and delays of outstanding inputs. Clearing the system incurs a fixed cost, and outstanding inputs are charged a delay penalty, which is a general increasing function of the quantities and delays of individual inputs. By recording the quantities and delays of outstanding inputs in a sequence, we model the clearing system as a tree-structured Markov decision process over both a finite and infinite horizon. We show that the optimal clearing policies, under realistic conditions, are of the on-off type or the threshold type. Based on the characterization of the optimal policies, we develop efficient algorithms to compute parameters of the optimal policies for such complex clearing systems for the first time. We conduct a numerical analysis on the impact of the nonlinear delay penalty cost function, the comparison of the optimal policy and the classical hybrid policy (ie, quantity and age thresholds), and the impact of the state of the input process. Our experiments demonstrate that (a) the classical linear approximation of the cost function can lead to significant performance differences; (b) the classical hybrid policy may perform poorly (as compared to the optimal policies); and (c) the consideration of the state of the input process makes significant improvement in system performance.