Reliability Prediction via Accelerated Testing with Imperfectly Controlled Conditions

The reliability of a highly reliable product is often estimated through accelerated testing with one or multiple stressors and well-designed stress profiles. Since the parameters of the product's life-stress relationship will change in response to stress variation, it is desirable to precisely control the designed testing conditions. However, widely used testing equipment, such as an environmental chamber, may not always meet such expectations with respect to the required level of accuracy. This may result in changes in the life-stress relationship during the test and, if ignored, potentially diminish the accuracy of reliability extrapolation at the use condition. In this paper, we propose a physics-informed statistical learning framework for product reliability prediction via accelerated testing with imperfectly controlled testing conditions. The proposed stress profile representation method and statistical estimation procedure partially relax the requirements for stringent control of applied stresses during accelerated testing. A dataset from a capacitor accelerated test involving both voltage and temperature stressors is modified and used to illustrate the proposed methodology. The results show that the proposed methodology is a useful tool for reliability prediction and is robust to moderate and continuous changes in accelerated testing conditions while requiring minimal added knowledge from the end user's perspective.