Recursive support vector censored regression for monitoring product quality based on degradation profiles

• Authors: Jong In Park;Myong K. Jeong
• Venue: Applied Intelligence
• Year: 2011

The time-consuming evaluation of a product's lifetime or quality often prevents manufacturers from meeting market requirements within the time allotted for product development. Degradation profiles obtained from harsh testing environments have been widely used in many applications to shorten the evaluation time. In this paper, we propose a novel recursive support vector censored regression (r-SVCR) technique to make a direct prediction on the lifetime based on the degradation profiles obtained in an accelerated testing setup. The proposed approach avoids potential bias introduced in the conventional prediction models due to accumulation of computational errors and misspecification of covariate models. Compared to standard support vector regression, our r-SVCR imposes the constraints on the derivatives of the regression function to ensure that the regression function is monotone over the input data range. Also, the r-SVCR accommodates the censored observations through our developed recursive estimation procedure, leading to error reduction. The hyperparameters of the proposed method are optimized based on the genetic algorithms (GAs).The proposed method represents a novel approach in that the functional form describing the degradation paths and even the relationship between input covariates and product degradation need not be specified. A real-life example of a degradation test in which both temperature and cut-off voltage stresses are employed to expedite a secondary rechargeable battery's failure during test intervals is presented to illustrate the proposed method and compare its performance with the conventional one. The results demonstrate the efficiency of the proposed method in predicting the lifetimes from the degradation profiles.