EPISTEMIC UNCERTAINTIES WHEN ESTIMATING COMPONENT FAILURE RATE

International Conference Nuclear Energy in Central Europe 2 Golf Hotel, Bled, Slovenia, September 11-14, 2 EPISTEMIC UNCERTAINTIES WHEN ESTIMATING COMPONENT FAILURE RATE Romana Jordan Cizelj, Borut Mavko, Ivo Kljenak Jožef Stefan Institute Reactor Engineering Division Jamova 39, 1 Ljubljana, Slovenia Romana.Jordan@ijs.si, Borut.Mavko@ijs.si, Ivo.Kljenak@ijs.si ABSTRACT A method for specific estimation of a component failure rate, based on specific quantitative and qualitative data other than component failures, was developed and is described in the proposed paper. The basis of the method is the Bayesian updating procedure. A prior distribution is selected from a generic database, whereas likelihood is built using fuzzy logic theory. With the proposed method, the component failure rate estimation is based on a much larger quantity of information compared to the presently used classical methods. Consequently, epistemic uncertainties, which are caused by lack of knowledge about a component or phenomenon are reduced. 1 INTRODUCTION Probabilistic Safety Assessment (PSA) is a mathematical tool, which may be used for systematical assessment of unavailability of complex systems, including components of nuclear power plants. Components, which are part of the analyzed system, are modeled with different probabilistic models. An exponential distribution is often selected for component models with a continuous random variable (failure time), whereas a binomial distribution is selected for component models with a discrete random variable (failure on demand). The necessary parameter for estimation of time- dependent component unavailability is the component failure rate λ. If enough raw data about component failures exist, the component failure rate can be estimated by the maximum likelihood method (equation(1)), or the Bayesian updating procedure (equation (2)) [1]: λ = n T (1) where n is the number of failures, and T the component operating time,

Romana Jordan Cizelj, page 2 of 9 p( λ E ) = p( λ ) L( E λ ) p( λ ) L( E λ ) dλ (2) where p(λ/e) is the conditional probability distribution of λ given evidence E, p(λ) the prior distribution, and L(E/λ) the likelihood function. If no raw numerical data about component failures exist, the value of the component failure rate should be selected from a generic database. Problems arising when using generic databases are: inconsistent definitions of component boundaries, unclear definitions of component failure modes, omitted data of component operating environment and operating conditions, differences of component designs, etc. These unknown parameters are sources of epistemic uncertainties, which are caused by lack of knowledge about a system, component, or phenomenon [2]. Thus, the use and the selection of a generic database are influenced by the analyst's subjective approach and by the completeness of the selected generic database. Apart from data about component failures, other information about component also exists and could be used for a specific estimation of component reliability/unavailability. Such information can be gathered for example from maintenance sheets and may be of quantitative or qualitative type. To avoid using data from generic databases and to decrease epistemic uncertainties, a method for specific estimation of a component failure rate, based on specific quantitative and qualitative data other than component failures, was developed [3]. The method is described in the proposed paper. 2 DESCRIPTION OF THE METHOD With the new method, a reliability parameter is determined by including more information explicitly into its calculation [4]. A formal mathematical procedure, which enables the calculation of a parameter, if the evidence about the parameter is increased, is the Bayesian learning procedure. A prior distribution, which contains data about the component design and material properties, is selected from a generic database. The likelihood function contains data about component operation, maintenance and surveillance. The exact connection between these additional data and component reliability parameter is often unknown. Furthermore, the value of data, which can be used for component reliability parameter estimation, is often not exactly known. Therefore, a likelihood function is built according to the principles of fuzzy set theory. The basic mathematical approach is shown in Figure 1. symptoms fuzzy inference system generic database likelihood L(E/ λ ) prior p( λ ) Bayesian inference posterior p( λ /E) Figure 1: Block diagram of main elements for estimation of a specific component failure rate

Romana Jordan Cizelj, page 3 of 9 First, for each component, a qualitative analysis of causes of component deterioration is performed. Then, the symptoms, which reflect component reliability, are defined. The influence of symptoms on component reliability is modeled using a fuzzy inference system: 1 1 2 K K if S is S j op S 2 is S j S is S j then RFR is RFR (3) i where S k is thelinguistic variable of the k-th symptom, S j k and RFR i are fuzzy variables, RFR (relative failure rate ) is a linguistic variable, K is the number of symptoms, and op is a logical operation, e.g. AND. With the linguistic variable RFR, the change of component reliability due to component operating conditions, maintenance and surveillance activities is described. The selection of its base variable depends on available information about the relation of component symptoms with component reliability. In the present paper, the base variable rfr of linguistic variable RFR is defined as the ratio of actual component failure rate λ a and component failure rate λ n, if the component is working in a normal operating environment: λa rfr = λ n (4) Fuzzy reasoning is performed following the Mamdani approach [5]. The resulting fuzzy set of the fuzzy reasoning is RFR r. Different interpretations of the fuzzy set RFR r are possible. In the present paper, the center-of-gravity (COG) procedure is used for calculation of crisp value rfr COG. Then, the interval of equally possible crisp values of rfr-s is determined [3]. If the result for a non-repairable component is the mean value of the component failure rate, it can be converted into mean time to failure (MTTF). The lower and the upper value of the interval resulting from defuzzification are then used for a further calculation (MTTF u, MTTF l ). The following exponential distribution can be used for the likelihood function [6]: L( E λ ) = e e λ MTTF l λ MTTF u (5) 3 ANALYSIS The proposed method was applied on an outdoor type, high voltage transformer with 4 MVA capacity, 4 kv high voltage winding, 21 kv low voltage winding and frequency 5 Hz. The transformer has a paper insulation, winding is cooled by forced oil circulation, while secondary cooling is performed by forced air. The following input data are available for the analysis: generic probability density distribution for component failure rate p(λ), and specific data about transformer hot-spot temperature, degree of insulation polymerization, ratio of hydrocarbon gases in oil, oil level, and acceptability of oil level and color of silicon-gel. In section 3.1, the building of the model is described, if the exact connection between the additional data and component reliability parameter is unknown. In section 3.2, the determination of input value for the fuzzy model is described, if the available raw data are not exactly known.

Romana Jordan Cizelj, page 4 of 9 3.1 Knowledge base development The knowledge base of a fuzzy inference system contains a list of fuzzy if-then rules and defines the membership functions of the fuzzy sets used in the fuzzy rules [7]. In the present paper, building of the knowledge base is described for the symptom 'hot spot temperature' [8]. The hot-spot temperature is the maximum temperature of the transformer windings and depends on ambient temperature, oil temperature, material properties of insulation, electrical loading and quality of oil (e.g. efficiency of cooling). The hot-spot temperature is only an estimation and not an exact calculation. It can not be measured directly, because the location of the hot-spot is not exactly known. The most common approach for modeling the relation between the hot-spot temperature and the transformer failure rate is to use the Montsinger calculation (6). A normal transformer deterioration occurs at temperature 98 C, and an increase (decrease) of temperature for every 6 C doubles (halves) the transformer relative aging rate[9]: λ λ whs 98 C = 2 whs 98 ( ) 6 = rfr (6) where rfr is the relative failure rate, λ 98 C the failure rate, if the transformer is operating at hotspot temperature 98 C, λ whs the failure rate, if the transformer is operating at hot-spot temperature whs, and whs the transformer hot-spot temperature in [ C]. The following assumptions were used when building the database of the symptom hotspot temperature: normal insulation aging occurs at hot-spot temperature 98 C with 2 C ambient temperature [1], the aging rate doubles for every 6 C increase in hot-spot temperature [1], the transformer MTTF at hot-spot temperature 98 C is 4 years [1]. The term set T WHS of the linguistic variable hot-spot temperature (WHS) is defined in equation (7), and is graphically represented in Figure 2. The base variable whs, corresponding to the linguistic variable WHS, is defined in equation (6). T WHS = 14 { approximately 8 or less,,approximately 98,,approximately or more} (7) The linguistic variable RFR WHS is defined as shown in Figure 3 and has the following term set: T = RFR WHS 128 { approximately. 125 or less,, approximately1,,approximately or more} (8)

Romana Jordan Cizelj, page 5 of 9 µ WHSi 1,5 74 8 86 92 98 14 11 116 122 128 134 14 146 whs [ C] Figure 2: Definition of linguistic variable hot-spot temperature (WHS) 1 µ RFR,WHSi,5,1 1 1 1 1 rfr Figure 3: Definition of linguistic variable relative failure rate WHS (RFR WHS ) With the definition of fuzzy if-then rules, the influence of hot-spot temperature on transformer failure rate is described. For example: if WHS is approximately 98 then RFR WHS is approximately 1 if WHS is approximately 14 then RFR WHS is approximately 2 3.2 Formation of input data The determination of input value for the fuzzy model is described for the symptom 'insulation polymerization' [11]. Paper insulation polymerization causes degradation of long cellulose molecules. The process depends on winding temperature, contents of water in paper, contents of oxygen in oil, insulation material properties, and quality of oil [1]. Measurements and deterministic calculations can provide an estimation of transformer degree of insulation polymerization and remaining life-time. Samples for measurements are taken arbitrarily at different places of the transformer insulation. Only a qualitative assessment can be made about the appropriate selection of the sample places. In addition, the initial value of paper polymerization is often unknown and a critical value of paper polymerization has not yet been commonly agreed [12]. Therefore we do not have accurate data about the degree of paper insulation polymerization (DP). The base variable wip of the linguistic variable insulation polymerization (WIP) is defined in equation (9). The base variable wip is defined as the ratio between normal transformer lifetime t n (= 4 years) and actual transformer lifetime given the measured DP.

Romana Jordan Cizelj, page 6 of 9 wip = t t n a (9) Results of measurements of paper polymerization are shown in Table 1. The measured data are different because samples are taken at different places of the transformer insulation and measurements were performed by two different independent groups. Table 1: Degree of insulation polymerization of the transformer Measurement groups Group 1 Group 2 Estimated lifetime Years of operation 15 17 15 17 Group 1 Group 2 342 31 2,7 19,7 357 34 21,2 2,6 353 35 43 325 * 19,4 Samples at different 6 494 632 68 22,5 42,5 places of transformer 43 45 23,9 24,7 insulation 361 44 21,3 24,3 355 47 445 18,6 36,6 635 62 577 72 41,4 * * unrealistic measurement results do not enable a meaningful estimation of transformer lifetime A transformer failure is often caused by a sudden mechanical or electrical stress and very rarely by a degradation process. Consequently, for estimation of transformer reliability, the critical places of transformer insulation with the lowest DP are relevant [13]. For further calculation, the lowest value of polymerization measured by each group was selected. Because of high uncertainties, introduced with the measurement and calculation of DP and remaining lifetime, the input value to the fuzzy inference system was built as a fuzzy set with the following lower and upper values (Figure 4): measurement of group 1: t a = 2,7 [years], wip = 1,9 measurement of group 2: t a = 19,4 [years], wip = 2,1 1,5 1,7 1,9 2,1 2,3 wip Figure 4: The input value of the symptom 'degree of paper insulation' Other input data for fuzzy inference system are as follows: hot-spot temperature: 99 C, color of silicon-gel: acceptable 98% of time, oil level: acceptable 1% of time, degree of insulation polymerization: from 1,9 to 2,1, furanic compounds in oil: (actual) failure rate is 1,3-times higher than expected (normal) failure rate, ratio of hydrocarbon gases: failure rate is higher than normal failure rate.

Romana Jordan Cizelj, page 7 of 9 The following prior distribution was selected for Bayesian analysis: lognormal distribution, mean = 2 1-6 [1/h], and EF = 1. 3.3 Results The result of estimation with fuzzy inference system is the fuzzy set relative failure rate r (RFR r ), which shows the possible values of the base variable rfr after evaluation of symptoms. The fuzzy set RFR r is shown in Figure 5. µ RFRr(rfr),8,6,4,2,7,5 2 4 6 8 1 rfr Figure 5: Resulting fuzzy set RFR r The prior, likelihood and posterior distribution function of Bayesian evaluation are shown in Figure 6. probability density,e+ 5,E-6 1,E-5 1,5E-5 2,E-5 λ p(. λ ) p( λ/ε ) mean L(E/.. λ ) Figure 6: Prior, likelihood and posterior distribution function of Bayesian evaluation The following results are compiled in Table 2: mean and median value of transformer failure rate, lower and upper value of 9% Bayesian reliability interval (λ 5, λ 95 ), and transformer mean time to failure (MTTF). Table 2: Results of Bayesian analysis λ 5 median mean λ 95 MTTF [years] prior 7,51E-8 7,51E-7 2,E-6 7,51E-6 57,1 posterior 3,E-7 2,5E-6 3,6E-6 9,28E-6 37,3

Romana Jordan Cizelj, page 8 of 9 The comparison of specific values with generic values shows important differences. In general, the lifetime of a nuclear power plant (NPP) is 4 years (with a trend in western type reactors to prolong it to 5 years). Comparison of NPP lifetime and specific MTTF shows that one can expect transformer failure before NPP end-of-life. Since a power transformer, which is installed at the output of the plant, was selected as an example, the decreased value of MTTF is a significant result and requires immediate actions. Possible measures can be improved inspection and maintenance, restoration of transformer normal operating conditions, as well as planing a replacement program of the transformer. 4 CONCLUSIONS A method is proposed, which enables the estimation of a specific value of a component reliability parameter, although no raw numerical data about the component failures exist. The method is a formal procedure that prescribes how the symptoms, which can indicate component reliability, should be treated for estimation of the component failure rate. With the proposed method, epistemic uncertainties, caused by lack of data, when no raw data about component failures exist, could be reduced. The method is general enough to be used for different components working in different operating environments. 5 ACKNOWLEDGMENTS The research was supported by the Ministry of Science and Technology of the Republic of Slovenia. 6 REFERENCES [1] IAEA. Procedures for conducting probabilistic safety assessments of nuclear power plants (level 1). Vienna, Austria: International Atomic Energy Agency; 1992 Jul. Safety series no. 5-P-4. [2] Helton, J. C. and D. E. Burmaster. Guest editorial: treatement of aleatory and epistemic uncertainty in performance assessments for complex systems. Reliability Engineering and System Safety. 1996; 54:91-94. [3] Jordan Cizelj, R.; B. Mavko, and I. Kljenak. Component reliability assessment using quantitative and qualitative data. Reliability Engineering & System Safety. Accepted for publication. [4] Jordan Cizelj, R. and I. Kljenak. A new approach for estimation of component failure rate. International Conference Nuclear Energy in Central Europe'99; Portorož, Slovenia. Ljubljana: Nuclear society of Slovenia; 1999: 35-312. ISBN: 961 627 13 X. [5] Bandemer, Hans and Gottwald, Siegfried. Fuzzy Sets, Fuzzy Logic, Fuzzy Methods with Applications. Great Britain: John Wiley & Sons Ltd.; 1995; ISBN: 471 95636 8. [6] Siu, N. O. and D. L. Kelly. Bayesian parameter estimation in probabilistic risk assessment. Reliability Engineering & System Safety. 1998; 62:89-116.

Romana Jordan Cizelj, page 9 of 9 [7] Jang, J.-S. R. ANFIS: Adaptive-Network-Based Fuzzy Inference System. IEEE Transactions on Systems, Man, and Cybernetics. 1993 May-1993 Jun 3; 23(4):665-685. [8] Jordan Cizelj, R. and B. Mavko. Assessment of the influence of operating conditions on component failure rate. European Safety and Reliability Conference ESREL 2; Edinburgh, UK. A. A. Balkema: 1533-1538. [9] Makuc, D.; K. Lenasi, and D. Velkov. Letna obtežna temperatura okolice in obremenitev transformatorjev/yearly weighted ambient temperature and transformer loading: 1-6. [1] Heathcote, Martin J. The J & P Transformer Book. Great Britain: Newnes; 1998; ISBN: 756 1158 8. [11] Jordan Cizelj, R. and I. Kljenak. The use of qualitative data for specific component failure rate estimation; Innsbruck, Austria. 2. [12] Babuder, M. and M. Končan-Gradnik. Primer uporabe kompleksne metode za ocenjevanje preostale življenjske dobe velikih energetskih transformatorjev. Tretja konferenca slovenskih elektroenergetikov CIGRE; Nova Gorica, Slovenia. Ljubljana: JBB d.o.o.: 12-41 - 12-48. ISBN: 961 928 8 1. [13] Babuder, Maks; Končan-Gradnik, Maja, and Miljavec, Damijan. Globalna ocena preostale življenjske dobe transformatorjev GT1 in GT2 v Nuklearni elektrarni Krško. Ljubljana, Slovenija: EIMV; 1997.