Optimal maintenance of a multi-unit system under dependencies
MetadataShow full item record
The availability, or reliability, of an engineering component greatly influences the operational cost and safety characteristics of a modern system over its life-cycle. Until recently, the reliance on past empirical data has been the industry-standard practice to develop maintenance policies that provide the minimum level of system reliability. Because such empirically-derived policies are vulnerable to unforeseen or fast-changing external factors, recent advancements in the study of topic on maintenance, which is known as optimal maintenance problem, has gained considerable interest as a legitimate area of research. An extensive body of applicable work is available, ranging from those concerned with identifying maintenance policies aimed at providing required system availability at minimum possible cost, to topics on imperfect maintenance of multi-unit system under dependencies. Nonetheless, these existing mathematical approaches to solve for optimal maintenance policies must be treated with caution when considered for broader applications, as they are accompanied by specialized treatments to ease the mathematical derivation of unknown functions in both objective function and constraint for a given optimal maintenance problem. These unknown functions are defined as reliability measures in this thesis, and theses measures (e.g., expected number of failures, system renewal cycle, expected system up time, etc.) do not often lend themselves to possess closed-form formulas. It is thus quite common to impose simplifying assumptions on input probability distributions of components' lifetime or repair policies. Simplifying the complex structure of a multi-unit system to a k-out-of-n system by neglecting any sources of dependencies is another commonly practiced technique intended to increase the mathematical tractability of a particular model. This dissertation presents a proposal for an alternative methodology to solve optimal maintenance problems by aiming to achieve the same end-goals as Reliability Centered Maintenance (RCM). RCM was first introduced to the aircraft industry in an attempt to bridge the gap between the empirically-driven and theory-driven approaches to establishing optimal maintenance policies. Under RCM, qualitative processes that enable the prioritizing of functions based on the criticality and influence would be combined with mathematical modeling to obtain the optimal maintenance policies. Where this thesis work deviates from RCM is its proposal to directly apply quantitative processes to model the reliability measures in optimal maintenance problem. First, Monte Carlo (MC) simulation, in conjunction with a pre-determined Design of Experiments (DOE) table, can be used as a numerical means of obtaining the corresponding discrete simulated outcomes of the reliability measures based on the combination of decision variables (e.g., periodic preventive maintenance interval, trigger age for opportunistic maintenance, etc.). These discrete simulation results can then be regressed as Response Surface Equations (RSEs) with respect to the decision variables. Such an approach to represent the reliability measures with continuous surrogate functions (i.e., the RSEs) not only enables the application of the numerical optimization technique to solve for optimal maintenance policies, but also obviates the need to make mathematical assumptions or impose over-simplifications on the structure of a multi-unit system for the sake of mathematical tractability. The applicability of the proposed methodology to a real-world optimal maintenance problem is showcased through its application to a Time Limited Dispatch (TLD) of Full Authority Digital Engine Control (FADEC) system. In broader terms, this proof-of-concept exercise can be described as a constrained optimization problem, whose objective is to identify the optimal system inspection interval that guarantees a certain level of availability for a multi-unit system. A variety of reputable numerical techniques were used to model the problem as accurately as possible, including algorithms for the MC simulation, imperfect maintenance model from quasi renewal processes, repair time simulation, and state transition rules. Variance Reduction Techniques (VRTs) were also used in an effort to enhance MC simulation efficiency. After accurate MC simulation results are obtained, the RSEs are generated based on the goodness-of-fit measure to yield as parsimonious model as possible to construct the optimization problem. Under the assumption of constant failure rate for lifetime distributions, the inspection interval from the proposed methodology was found to be consistent with the one from the common approach used in industry that leverages Continuous Time Markov Chain (CTMC). While the latter does not consider maintenance cost settings, the proposed methodology enables an operator to consider different types of maintenance cost settings, e.g., inspection cost, system corrective maintenance cost, etc., to result in more flexible maintenance policies. When the proposed methodology was applied to the same TLD of FADEC example, but under the more generalized assumption of strictly Increasing Failure Rate (IFR) for lifetime distribution, it was shown to successfully capture component wear-out, as well as the economic dependencies among the system components.