Reliability prediction using support vector regression

Abstract

Reliability prediction of machinery is crucial to schedule overhauls, spare parts replacement and maintenance interventions. Many AI tools have been used in order to provide these predictions for the industry. Support vector regression (SVR) is a nonlinear regression technique extended from support vector machine. SVR can fit data flexibly and it has a wide variety of applications. This paper utilizes SVR combining time series to predict the next failure time based on historical failure data. To solve the parameter selection problem a method has been proposed. This method approximates the widely used leave-one-out method. To bound the prediction error, a confidence interval is proposed based on the non-homogeneous poisson process. A numerical case from the mining industry is presented to demonstrate the developed approach.

Appendix

Assuming that one failure occurs at time t _i , the probability of at least one failure occurring during time t to time t _i+1 is:

$$ P[N(t) \ge 1] = 1 - P[N(t) = 0] $$

(A.1)

where N(t) denotes the number of failures between t and t _i . As the number of failures in the interval [t, t _i+1] is Poisson-distributed, by using the Poisson theorem:

$$ P[N(t) = 0] = e^{{ - \int_{{t_{i} }}^{{t_{i + 1} }} {\lambda (t)dt} }} $$

(A.2)

Therefore, the probability of failure during [t, t _i+1] is:

$$ P[t \ge t_{i} ] = 1 - e^{{ - \int_{{t_{i} }}^{{t_{i + 1} }} {\lambda (t)dt} }} $$

(A.3)