Online Remaining Useful Lifetime Prediction Using Support Vector Regression

An accurate prediction of remaining useful lifetime (RUL) in high reliability and safety electronic systems is required due to its wide use in industrial applications. In this paper, we propose a novel methodology for online RUL prediction, using support vector regression (SVR) model. Through Cadence simulations with 22nm CMOS technology library, we demonstrate that frequency degradation follows a trackable path and depends on temperature, voltage and aging. This characteristic is exploited for training the SVR model, validated over 20 years of aging degradation. Our methodology is capable of highly accurate RUL estimation, requiring a ring oscillator (RO), temperature sensor and trained SVR software model. Using a supply voltage of 0.9 V and variation in temperature from 0 degrees C to 100 degrees C, 13 and 21 stage RO show 90 percent cases with a RUL prediction deviation of +/- 0.2 years, and the remaining between +/- 0.75 and +/- 0.8 years, respectively. Furthermore, with voltage variation from 0.7 to 0.9V, with steps of 0.05V and four representative temperatures (25, 50, 75 and 100 degrees C), the 13-RO shows 52 percent cases between +/- 0.2 years, 21-RO has 80.5 percent cases concentrated between +/- 0.2 years of RUL prediction deviation and remaining cases for both ROs are located between +/- 0.8 years.

Automated summary

This paper proposes a novel methodology for online prediction of remaining useful lifetime (RUL) in high reliability and safety electronic systems using support vector regression (SVR) model. The methodology utilizes frequency degradation as a trackable path and depends on temperature, voltage, and aging. Results show that the methodology can accurately estimate RUL with a high level of accuracy.