Research On Methdos For Condition Based Product Residual Life Prediction And Predictive Maintenance Scheduling

Condition-based maintenance (CBM) is an emerging reaseach topic with vitalapplication and economic significance. With CBM being the background, thisthesis develops and extends several methods for prognostics (i.e. equipmentresidual life (RL) prediction and predictive maintenance (PdM) scheduling)through theoretical investigation, simulation, and experimaetal validation.Equipment RL prediction and PdM have essential impact on the effectivenessof a maintenance policy but have not been fully investiagated in literature.First of all, the thesis studies two types of methods for equipment RLprediction. Two studies are devoted to the topic of similarity-based residuallife prediction approaches. Within the framework of similarity-based RLprediction methods, a generalized similarity measure is proposed. By theproposed similarity measure, more weight is assigned to a system’s mostrecent performance than to its former performance when measuring itssimilarity with other systems. The experimental results from a numericalexperiment and a case study on RL prediction of ball grid array (BGA)packages show that, compared with the predictions with the classicalsimilarity measure, statistically more accurate predictions can be achievedwith the proposed similarity measure. A framework of similarity-basedresidual life prediction approaches is proposed, where historical samples thatfail and do not fail (do to preventiven maintenance or suspension) are bothutilized. Within the framework, two solutions (namely solution A and solutionB) are proposed to estimate the lifetimes of the preventively maintained orsuspended historical samples, and to utilize their degradation histories. Anextensive numerical investigation verifies the superiority of the proposedframework (using solution A) over the corresponding classical SbRLPapproach from the case of limited failed historical samples to the case ofabundant failed historical samples. In addition, the investigation results revealthat the proposed framework (using solution B) is ineffective when failedhistorical samples are limited, but its performance improves fast with the increment of available failed historical samples. Afterwards, two studies aredevoted to the topic of proportional harards model (PHM) based equipmentresidual life prediction:1) A Two-zone PHM is proposed based on the factthat many systems’ degradation processes can be devided into a stable zoneand an unstable zone;2) An extended PHM (EPHM) is proposed for RLprediction of systems receiving partial recovery preventive maintenance (PM)acts (i.e. imperfect PM acts), while the classical PHM is not applicable tonon-repairable systems. The results from numberical experiments show that:1)For systems whose degradation processes can be devided into a stable zoneand an unstable zone, the proposed two-zone PHM provides statistically moreaccurate and reliable predictions close to system failure, compared with theclassical PHM;2) For systems receiving imperfect PM acts, the proposedEPHM provides statistically more accurate and reliable predictions close tosystem failure, compared with the classical PHM.Secondly, the thesis contributes three studies on PdM schedulingmodels/policies. Compared with the classical imperfect PM schedulingmodels, an updated sequential PdM (USPdM) policy is proposed consideringthe effect of imperfect PM acts. The USPdM calculates and updates theoptimal PM schedules for continuously monitored single-unit degradingsystems to minimize the long-term maintenance cost rate. A case study onPdM scheduling of wearing drill bits show that:1) The proposed USPdMmodel yields PM schedules that are consistent with the change in systemstates and2) the USPdM is able to quickly react to drastic degradation of thesystem and provide an optimal PM schedule in real time. In practice, itusually takes more time to restore a much degraded system to a desire statethan to restore a slightly degradaed system to the same state. Based on suchfact, a PdM policy is proposed for maximizing the average system availabilityof continuously monitored single-unit systems, considering the effect ofdegradation correlated maintainability. In the PdM policy, the maintainabilityof PM acts is correated with the accumulated degradation level, which ismodeled by the recently proposed proportional repair model (PRM). Theresults from the numerical experiment show that, for a batch of systemswhose PM maintainability is indeed correlated with the accumulateddegradation level, simply neglecting such effect would lead to suboptimal PMschedules. In addition, the results provide evidence of the superiority of thePdM policy over the corresponding PM polices in terms of higher averagesystem availability. The3rdstudy on PdM scheduling models/policies proposes a modularized framework of PdM scheduling. Based on themodularization treatment, a PdM scheduling model can be established byintegrating the classical PM/CBM scheduling models with components’failure probability estimated based on their degradation variables, and both ofthem have been widely studied. An unreported PdM scheduling model forserial systems is further established. Numerial experiments are conducted tocompare the optimal PM schedules from the PdM scheduling model for serialsystems and the counterparts from the corresponding PM scheduling model,and investigate the characteristics of the PM schedules from these two models.The experimental results show that the PdM model is more effective inmaintenance cost reduction. The results also highlight the individual-oriented,dynamically updating characteristics of the PM schedules from a PdMscheduling model, which is different from the population-oriented, static PMschedules from a PM scheduling model.Thirdly, the thesis studies two integration schemes of two main aspects ofprognostics (i.e., equipment RL prediction methods and PdM schedulingmodels,), which has not been reported in literature. In the1ststudy, astatistically planned and individually improved (SPII) PdM policy is proposedfor continuously monitored single-unit systems. The SPII PdM policysimultaneously takes advantage of1) the capability of the classical statisticallifetime distribution based PM scheduling policy in long-term planning and2)the capability of RL prediction methods in improving individual performance.The value of the SPII PdM policy is that it demonstrates the possibility ofpartially applying the emerging RL prediction techniques in the widely usedstatistical lifetime distribution based PM policy in an (approximately)theoretically effective manner. By reasonably applying the RL predictiontechniques to a part of (but not all) individuals, a better performance of themaintenance policy can be expected, while the performance of the classicalPM policy can first provide baseline indexes for long-term palnning prior tosystem operation or real production. In the2ndstudy, a PdM framework isproposed for continuously monitored single-unit systems, which integratesequipment RL prediction methods and PdM scheduling models. Within theproposed framework, PdM scheduling is trigrred only when the RLpredication methods demonstrate that the system is close to failure. A lifetimemargine is proposed to infer that most systems do not fail before PdMscheduling is triggered. A numerial experiment demonstrates theimplementation procedures of the proposed framework. In the numrial experiment, when the noise level of the degradation process is relatively low,compared with the classical PdM scheduling model which starts maintenancescheduling once a system bengins operating, the proposed PdM framework issimilar effective in failure prevention and more economical in PM costsaving.Finally, the thesis conducts two application studies on condition monitoringdata. Based on the condition monitoring and event data, the whole processfrom degradation process analysis to determination of PM schedules isdiscussed. A quantitative comparison among a preventive replacementscheduling model using condition monitoring data, a classical preventivereplacement scheduling model using statistical distribution of entire lifetimesof BGA packages, and a PdM scheduling model. Based on theimplementation results, the reason for maintenance performance improvementdue to the application of codition monitoring data is investigated. Theinvestigation results demonstrate the relationship bewtten the performance ofa maintenance policy and the accuracy and reliability of lifetime predictionwhen determing PM schedules. As the degradation model of the degradationindicator provides the most accurate and reliable lifetime prediction of BGApackages, the PdM scheduling model gives the best performance. A on-oribtreliability prediction method for spacebrone effective load is proposed to fullyutilized the remote monitoring variables, which has already been set on manykey circuits of spacebrone main load but has not been fully utilized. A casestudy is provided to illusate the application process of the proposed method.