AspenTech blogger Mike Brooks has provided a critique of ‘remaining useful life’ (RUL) as a metric for predictive maintenance applications. While RUL is a commonly accepted KPI, Brooks claims that it is a ‘flawed concept’. Traditional maintenance is organized around statistical assumptions of failure mechanisms, mostly related to normal wear and tear. However, studies by ARC Advisory Group have shown that over 80% of asset failures are caused by ‘errant process’ rather than age-related issues, making conventional preventive maintenance strategies ineffective. Examples of such errant processes include operator error, lightning strikes and voltage spikes or episodic cavitation in a pump. Most of these random events make the reliable prediction of RUL impossible. For more on AspenTech’s approach to prescriptive analytics and equipment failure read the white paper on ‘Low-touch’ machine learning for asset management.
In a MathWorks article, Aditya Baru has no qualms about using RUL as a KPI and suggests three ways of estimate RUL for predictive maintenance. MathWorks’ Predictive maintenance toolbox computes RUL depending on what data is available - lifetime machine data, run-to-failure histories and threshold values of know failure condition indicators. Computing RUL may leverage proportional hazard models and probability distributions of component failure times. For example, a battery’s discharge time may be estimated from past discharge rates and covariates, such as operating temperature and load. The MathWorks approach involves physics-based forward modeling and Kalman filtering. More RUL examples here.