A Distribution-Level Combinational Model to Improve Reliability Prediction Accuracy

Abstract

Single reliability growth model usually only captures partial knowledge of a failure process. A combinational model tries to capture more knowledge by integrating two or more reliability growth models. Unlike the existing linear combinational models that simply adds up the weighted results by G-O, M-O and L-V model, this paper proposes the combinational model from G-O and S-Shaped model, at the level of failure distribution to reduce fitting errors and to maintain the mathematical properties of non-homogenous Poisson process. To evaluate the effectiveness of the proposed model, we use the failure data sets (21 projects) available in public sources. Ten out of the twenty-one projects, which pass the distribution test and have feasible solutions in parameter estimation, are selected to conduct experiments. We use mean squared error (MSE) to evaluate the historical predictive validity. The results show that our model is consistently stable and has lower MSE. It reduces 51.3% MSE of G-O, 67.2% MSE of S-Shaped, and over 56% MSE of the three linear combinational models in average. The proposed model tends to have a larger estimation of the expected number of failures, which can overcome the under estimation by G-O and S-Shaped model in some degree.