[Paper Review] On-line Survival Analysis of Power Electronic Converters Using Step Noise-Cox Processes

Pourya Shamsi|arXiv (Cornell University)|Jul 10, 2015

Electromagnetic Compatibility and Noise Suppression23 references3 citations

TL;DR

This paper proposes an online survival analysis method for power electronic converters using step noise-Cox processes to estimate Mean Residual Life (MRL) under dynamic stress levels. By modeling time-varying stress as a stochastic process and applying a doubly stochastic Poisson process (Cox process), the method enables real-time, adaptive MRL estimation that outperforms traditional off-line models by incorporating live stress measurements, reducing aging overestimation under variable operating conditions.

ABSTRACT

This paper is focused on survival analysis of electrical components. The main goal of this paper is to develop a method for on-line estimation of the Mean Time To Failure (MTTF) of electrical components under dynamic stress levels. The proposed method models the variations of the stress levels as a stochastic process. Hence, a stochastic failure rate function can be developed for each electrical component. Later, this function is used as the underlying rate of a doubly stochastic Poisson process (known as Cox processes). Furthermore, this Cox process is used for on-line estimation of the Mean Residual Life (MRL) using the observed stress levels. The proposed method provides a good estimate of the age and life expectancy of each component. An experimental case study is provided to demonstrate the proposed method.

Motivation & Objective

• To develop an on-line method for estimating Mean Time To Failure (MTTF) and Mean Residual Life (MRL) of electrical components under time-varying stress levels.
• To address the limitations of conventional off-line reliability models that assume constant worst-case stress levels, leading to overestimation of aging.
• To model the stochastic variation of operational stress (e.g., voltage) as a step-noise process to better reflect real-world dynamic conditions in systems like microgrids and electric vehicles.
• To improve accuracy in age and life expectancy estimation by updating failure rate functions in real time based on observed stress data.

Proposed method

• Model the time-varying stress levels (e.g., battery voltage) as a step-noise stochastic process to capture dynamic operational conditions.
• Use the observed stress levels to define a stochastic failure rate function, which drives a doubly stochastic Poisson process (Cox process) for failure time modeling.
• Estimate the Mean Residual Life (MRL) on-line by conditioning the Cox process on the cumulative history of stress measurements up to each time point.
• Incorporate a Weibull-based bathtub failure rate function to model time-dependent failure behavior, including infant mortality and wear-out phases.
• Use moving average or cumulative mean of observed voltages to estimate the expected future stress level (E_G[ν]) for predicting future failure rates.
• Apply clamping to MRL estimates to reflect physical limits, such as a maximum MRL of 15,500 hours due to mechanical/chemical wear-out.

Experimental results

Research questions

• RQ1 How can the Mean Residual Life (MRL) of a power electronic component be estimated in real time under dynamically varying stress levels?
• RQ2 To what extent does using a stochastic failure rate function based on live stress measurements improve MRL estimation accuracy compared to static, worst-case models?
• RQ3 How do different estimation strategies for future stress levels (e.g., last observed value vs. cumulative mean) affect the responsiveness and accuracy of the on-line MRL prediction?
• RQ4 Can the Cox process framework effectively model the time-to-failure distribution of components under non-homogeneous stress conditions in real-world applications?

Key findings

• The proposed on-line method reduces overestimation of aging by adapting the failure rate to actual measured voltage levels, unlike off-line models that assume constant 120V stress.
• The average aging factor for the capacitor was estimated at 3.7, indicating the component ages 3.7 times faster than calendar time, but this factor is updated dynamically based on real-time voltage data.
• Model 1 (using last observed voltage) showed the fastest response to step changes in voltage, while Model 3 (using cumulative mean and base aging factor) had the slowest but more stable response.
• All proposed models converged to the same long-term MRL estimate as t→∞, confirming consistency in steady-state behavior.
• The infant mortality phase ends at approximately 2,500 hours, as shown by the MRL curve, indicating the model captures early failure dynamics.
• The on-line MRL estimation method achieved a more accurate representation of component age and residual life by dynamically adjusting the failure rate function based on observed stress history.