Reliability Engineering and System Safety 24 (1989) 21-32 A Maintenance Model for Repairable Systems L. C. Tang & F. O. Olorunniwo* Department of Industrial & Systems Engineering, National University of Singapore, Kent Ridge, Singapore 0511, Singapore (Received 12 March 1988, accepted 1 June 1988) ABSTRACT A model using a sequence of distributions, with dependent parameters, for a repairable system is proposed. By using a set of failure data containing successive times to failures, the parameters of the model are estimated and functional forms relating the parameters to the state of the system are obtained. Simulation is then conducted to determine the optimal age maintenance scheme for such a system. Finally, a possible extension is suggested by introducing stabilizing rate and deterioration rate of a class of repairable systems so that a stochastic model for such repairable systems can be fully determined. INTRODUCTION For years, confusion in assessing the reliability of repairable systems has pervaded. This is because the tacit assumption that times to failure are statistically independent and identically distributed is applied to all problems without knowing that it is only applicable to a non-repairable item. Ascher! attempted to remedy this by using stochastic processes for a reliability model but his work initially received very little attention. It was not until the work by Ascher and Feingold,? which is a compilation of the work they have done on repairable systems since 1968, that the analysis of times between failures for a repairable system gradually becomes the subject of interest in the reliability community. * Currently on a visiting appointment from the Department of Industrial Engineering, University of Ibadan, Nigeria 21 Reliability Engineering and System Safety 0951-8320/89/$03'50 © 1989 Elsevier Science Publishers Ltd, England. Printed in Great Britain2 LC. Tang, F. O. Olorunniwo The common misconception is the use of a single distribution function for the times to failures throughout the life of the system. This is not correct since the changes in interarrival times to failures due to deterioration or improvement cannot be represented by a single distribution function. In fact, each successive time to failure follows a failure distribution which may be different from the others. The collection of these distributions constitutes a stochastic point process for a repairable system. So far, much of the recent work concentrates on models using Nonhomogeneous Poisson Processes (NHPP) such as Ascher and Feingold? and Engelhardt and Bain,’ which may appear too mathematical for some practitioners, Moreover, as the analysis is always distribution free, it could be difficult to visualize. In this paper, a methodology for modelling a deteriorating repairable system using practical data, with a collection of distributions of successive times to failures is presented. By performing a regression on sample data as an example, we demonstrate the dependency between the distributions and show how the process whereby the MTBF is decreasing can be appreciated. In addition, Drenick* limit theorem, which indicates that the failure rate of the repairable system will tend to a constant after a sufficiently long time due to random age mix, can also be physically visualized. This model is also applied to determine the optimal age maintenance policy of a repairable system by simulation, A possible extension to the model is made by introducing two new terms, ie. stabilizing rate and deteriorating rate. The former is a measure of how soon the failure rate approaches constant and the latter determines how fast the decrease in mean time to successive failures will be. With these two values specified, it will be shown that the stochastic model for a repairable system (deteriorating or improving) can be fully determined. A SIMULATION MODEL FOR THE MAINTENANCE OF A REPAIRABLE SYSTEM In this section we discuss a simulation model for determining the maintenance policy for a repairable system. We shall first discuss the assumptions we make and then give some general considerations applicable to repairable systems in general and to our model in particular. Definitions and assumptions The model developed in this paper applies to a major failure and hence the ‘failure rate’ is referring to the failure rate of a major failure. The terms used in this paper are defined as follows:A maintenance model for repairable systems 23 1. State i: The situation where (i— 1) failure has taken place and ith failure is impending, 2, Preventive Maintenance (PM }: Maintenance action which checks for and rectifies flaws or wear-out signals that will lead to a major failure. After PM, all the potential causes of major failure are removed and the associated failure rate is restored to the initial value of the state (cf. ‘as-good-as-new’ assumption). 3. Breakdown Maintenance (BM): Maintenance action which is carried out when there is a major failure. When a major breakdown occurs, the system transits from state /to state (i + 1) and this triggers the use of another failure distribution which is determined from the next successive times to failures; so that the actual effect of BM is reflected on the following failure distribution. As such the assumption on after-repair-effect is avoided in that failure is treated separately. 4, _ Interarrival time: The times to failures measured from BM or PM. To take other minor failures into consideration, corrective maintenance is assumed to take place in rectifying such a fault with minimal repair such that the failure rate of the major failure is not affected at all. The cost of corrective maintenance is assumed to be negligible compared to the costs of BM and PM. Simulating the maintenance activity With the above definitions and assumptions, age replacement policy is implemented with time to replace either equal to the interarrival time in state ior the PM cycle, whichever comes first. This policy is well-known in reliability discipline. However, most work in this area such as Glasser’ and Kabir,® always assume the existence of a global failure distribution. Hence they are really meant for a non-repairable system. In the present case, the model for a repairable system consists of a sequence of distributions. Obviously, the optimal age maintenance cycle associated with each distribution could be obtained from graphs such as those provided by Kabir.® However, if the management decides to carry out age maintenance at equal intervals 1, over a planning horizon, an analytical optimal solution may not be so obvious. In addition, the effect of PM is difficult to predict due to human factors such as incomplete repairs, learning curve, etc. Hence, Monte Carlo Simulation is resorted to so that the occurrence of the human factors is modelled by the randomness introduced in the simulation through the use of the random numbers generated. The planning horizon In practice, a maintenance scheme will be reviewed after a pre-assigned planning horizon. However, most theoretical analysis such as Barlow and_ L. ©. Tang, F. 0. Olorunnivo Hunter’ and Lie and Chun’ always assume an infinite planning horizon for mathematical tractability. In the present simulation model, the planning horizon is fixed by two criteria, i.e. a prescribed total mileage travelled and a pre-determined number of failure, whichever comes first. We relate the total mileage, Ty,» to the allowable number of failure, in. by the following expression: Trnax= y MTBF, a) = where MTBF, is the mean mileage between the (i — 1)th and ith failures. This is essential as the total mileage travelled, under different ,, to reach a specific failure number will be significantly different. Thus to compare the performance of different 7, on the same basis, the total mean interarrival mileage up to a specific failure number must also be a terminating criterion. THE MODEL Ascher and Feingold, and Tullier® have shown, from data of repairable systems in general, that times between failures tend to decrease, and the distribution of times between failures (i— 1) and i are essentially different. The data used by Ascher and Feingold? are shown in Fig. 1 and are for the number of miles between successive major failures of bus engines. It can be observed that the distribution of time between ith and (+ 1)th failure appears normal for i= 1, and approaches exponential as i— 5. In our model we shall assume that each of these distributions is Weibull, although we recommend, as we demonstrate in our example problem, that the actual distribution be determined by the goodness of fit method. Since the Weibull distribution can apply to several real situations, we adopt that distribution here. Letting f,, 0, be the shape and scale parameters of the Weibull distribution for the system in state i, we need to establish the dependency or otherwise of B,oniand any other parameter. That is, there may be functions /{-) and g(-) such that B,=fli, 1) 9; = ali, ¥2) where is a collection of some other parameters. As will be shown later, ft-) and g(-) for our example are only functions of i. The approach we shall take therefore is to determine f,, @,, i= 1,2,... byA maintenance model for repairable systems 25 a 4 181 FAILURES WOK 200K 250K % 0 | 105 FAILURES SOK 100K BOK 200k 750K ye UM@ER OF FAILURES IN 10K MILES Fig. 1a. Interarrival miles of first and second failures of bus engines (191 engines originally put into service). (Source: Ascher and Feingold,’ p. 87.) Note: 1 mile = 1-6093 km, 20 % = voy ealunes 5 250K = go 2 M4 : 2 7 % = wl ah 96 Fates : 750K ah, — % 0 4 ales, sox Tom OKOKER he 1b. Interarrival miles miles of third, fourth and fifth failures of bus engines (191 engines originally put into service). (Source: Ascher and Feingold,? p. 87.) Note: 1 mile = 1-6093 km.26 L.C. Tang, F. O. Olorunnivo fitting a Weibull distribution to the data of the ith failure state and use a regression analysis to determine fli, y,) and g(i, 7). As the model uses a collection of Weibull distributions, each having its parameters related to the state of the system, the complementary cumulative distribution function can be written as R(t) =exp [ . ("| ) When k; PM has been carried out, and with the assumptions stated earlier, the failure rate is given by AMG)= At Kil) — Kily $4 kjf,, k = 0, 1,2,.... the upper bound of R¥(r,) is given by exp[—A(i,/0,)"1. Hence, fi 0< RNt) exp[—k;,(t,/0,"'] after PM wouldA maintenance model for repairable systems 29 imply that there is a reliability growth. This is contradictory to the intended assumption of ‘as-good-as-new’ in the literature. AN EXAMPLE APPLICATION The example application provided below serves to illustrate how the proposed algorithm can be used in a real life situation. The data used, although taken from the literature, are real data. The reader is cautioned that the features characterising the data as discussed later, may not be replicated exactly in all situations. Our algorithm however, can always be applied. Failure data and their distributions Ascher and Feingold? used interarrival mileage between major failure of bus engines to demonstrate the effect of successive repairs on the reliability of a deteriorating system. The data, which consist of failure distributions of the interarrival mileage up to the fifth failure, are shown in Fig. 1. These data are grouped in an interval of 10000 miles (16 093 kilometers). To derive a better distribution for each interarrival mileage and to estimate the respective parameters, the number of failures in each interval are read from the bar charts and the same number of uniform variate bounded by the interval are generated. This is to facilitate the Kolmogorov-Smirnov (K-S) test for goodness of fit. Each bar chart (grouped data) is transformed into five sets of continuous data in order to average out the error, and then fitted to different distributions. From Fig. 1, it can be seen that the distributien of the first failure resembles a Normal Distribution and there is a trend of approaching Exponential Distribution for the subsequent failures. Hence, distributions like Weibull, Gamma and Lognormal which have this property with changing parameters are fitted to the data. Normal and Exponential Distributions are also used for confirmation and comparison purposes. As data for each interarrival mileage (i.e. each bar chart) are generated five times, five sets of results are obtained for each interarrival mileage of the ith failure. It is found that only Normal Distribution provides the best fit to the first failure which satisfies the K-S test at the 5% significance level. However, both Weibull and Gamma Distributions also give a satisfactory fit at the 1% significance level of the K-S test. Moreover, these two distributions give a consistently good fit, at 5% significance level of the K-S test, to the subsequent failures data. We have therefore chosen the Weibull distribution to model each failure. The average values of the parameters for these two30 L. ©. Tang, FO. Olorunniwo TABLE 1 Parameters of the Failure Distributions ith failure Gamma distribution state ee Weibull distribution Shape parameter Scale parameter Shape parameter Scale parameter 6596 14-69 1918 11504 1 2 2368 29-62 1-146 80-77 3 1879 28-61 ri 59:30 4 1810 22:90 1-043 46-12 5 1-616 20:30 1-026 36-46 distributions are computed and are shown in Table1 for comparison purposes. We shall use the values for the Weibull distribution for our further analysis. Dependency of parameters In Table 1, the parameters exhibit an obvious trend in relation to the number of failures that have taken place, i.e. the state of the system. Hence, regression analysis is carried out, using SAS, to explore the potential functional form. As suggested from real life examples and Drenick* theorem, that after sufficiently long age, the occurrence of failure will tend to a constant value, an exponential decay for the values of (B — 1) and (@ — @,,i,) with respect to the state iis assumed. Strikingly, the assumption is found to be tallied with the estimated values of f, and 0,. The following expressions (eqns (7) and (8)) giving r?>=090 and r?>=0-99 respectively where r is the correlation coefficient, are found for f(-) and g(-): = 1+ 1358 exp(—0835 x i) (7) 46-5 exp (—0-286 x i) (8) where , and 6, are the shape and scale parameters of the Weibull distributions in state i. Note that 0,,, = 0 could be due to the strictly deteriorating nature of the present system (i.e. the truck engine) or insufficient data for subsequent failures. However, since only major failures are considered, the maximum i would be quite small and 9, would not drop to an unrealistic value. Simulation result for example problem When a brand new truck engine belonging to the same population as those given in the data is put into operation, the optimal preventive maintenanceA maintenance model for repairable systems 31 cycle, ¢*, can be determined by the algorithm if the management decides to service it at equal intervals. As the distributions up to the fifth failure are available, i,,,, is set equal to 5 and the corresponding Tua, 300 kilomiles by eqn (1). Sample simulation results give 1* = 30000 miles (48 280 kilometers) for cost ratio CR (= cost of BM/cost of PM) of 10, and r* = 20000 miles (32 186 kilometers) for CR = 5. A POSSIBLE EXTENSION Practical experience in operating repairable systems has supported, as discussed earlier, the fact that the distributions of the time to failure for successive failures are different. By assuming that each of these distributions is Weibull, we have used the example problem here to demonstrate the dependency of these parameters on the number of previous failures. Although the functional relations 8, = fli) and 6, = g(i) are found to be exponential in our particular example, these relationships may differ for different systems, the exact expressions being obtainable using the procedure outlined earlier in this paper. Assuming that the functional relationships (7) and (8) hold for a particular situation, it is interesting to characterise the deterioration behaviour of such arepairable system. In eqn (7), the coefficient in the exponential operator, ic. —0-835, can be interpreted as the coefficient of approaching steady state where f = 1. The stabilizing rate, r,, can be defined as B=NYeae cen B—Vaaes which determines the rate of f approaching 1. With this, the present r, for the truck engine is r, (9) xp [—0'835(i + 1) + 08351] =exp[—0:835] 3.4% Similarly in eqn (8), —0-286 can be interpreted as the coefficient of deterioration, and the deterioration rate, r,, can be defined as — Sates 6 Oat Present deterioration rate is then given as exp (—0-286) 751% ra L. ©. Tang, FO. Olorunniwo For a class of repairable systems characterised by a behaviour similar to that of the truck engines discussed here, with these two rates assumed from experience, and the parameters of Weibull distribution for first failure estimated from life test, the stochastic model of such a repairable system is complete as the distribution at each state can then be determined. As Tullier® observed, there may be different trends of f and @ in the successive failure distributions, as was his experience with shipboard machinery. Such data may not exactly fit our postulated extension, but the procedure outlined earlier can still be used. CONCLUSIONS In this paper, a model for a repairable system is proposed and applied toa set of real life data. This model, which consists of a sequence of Weibull Distributions for successive times to failures, will be more appealing, than other NHPP, to practitioners in the reliability community. Through simulation, the optimal age maintenance cycle is determined. An extension to the model using new parameters, i.e. stabilizing rate and deterioration rate, is also postulated for a certain class of repairable systems. REFERENCES Evaluation of repairable system reliability using the ‘bad-as-old’ Trans. on Reliability, R-17(2) (1968) 103-110. 2. Ascher, H. E. & Feingold, H., Repairable Systems Reliability. Marcel Dekker. Inc., New York, 1984, 3. Engelhardt, M. E. & Bain, L. J., On the mean time between failure for repairable systems, JEEE Trans. on Reliability, R-35(4) (1986) 419-422, 4. Drenick, R. F., The failure law of complex equipment. J. Society of Industrial & Applied Mathematics, 8 (1960) 680-690. 5. Glasser, G. J,, Planned replacement: some theory and its applications, J. Quality Technology, 1 (1969) 110-119 6. Kabir Zohrul, A. B. M., A new graphic presentation of cost of optimal replacement policies. Reliability Engineering, 17 (1987) 57-71. 7. Barlow, R. E. & Hunter, L., Optimum preventive maintenance policies. Operations Research, 8 (1960) 90-100. 8. Lie, C. H. & Chun, Y. H., An algorithm for preventive maintenance policy. IEEE Trans. on Reliability, R-35 (1) (1986) 71-75. 9. Tullier, P., Determining reliability and degradation of shipboard machinery. Proceedings of ARMS (1976) 239-244. 10. Malik, A. K. M., Reliable preventive maintenance scheduling. A//E Transactions, 11(3) (1979) 221-228. Il. Malik, A. K. M., Equipment replacement and its real causes. Maintenance Management International, 5 (1985) 51-61.