Feature Selection for Remaining Useful Life Prediction
For reliable remaining useful life (RUL) estimations, you want a condition indicator whose change over time is observable and connected with the system degradation process in a reliable, measurable way. The remaining useful life of a machine is the expected life or usage time remaining before the machine requires repair or replacement. Predicting remaining useful life from system data is a central goal of predictive-maintenance algorithms.
After you identify condition indicators (see Condition Indicators for Monitoring, Fault Detection, and Prediction), selecting useful condition indicators out of all available features is the next step in building a reliable RUL prediction model.
Predictive Maintenance Toolbox™ offers three feature selection metrics for accurate RUL prediction: monotonicity, trendability, and prognosability. These metrics rank the identified condition indicators on a scale ranging from 0 through 1. A higher ranked feature tracks the degradation process more reliably and hence, is more desirable to train the RUL prediction model.
Monotonicity characterizes the trend of a feature as the system evolves toward failure. As a system gets progressively closer to failure, a suitable condition indicator has a monotonic positive or negative trend. For more information, see
Prognosability is a measure of the variability of a feature at failure relative to the range between its initial and final values. A more prognosable feature has less variation at failure relative to the range between its initial and final values. For more information, see
Trendability provides a measure of similarity between the trajectories of a feature measured in multiple run-to-failure experiments. Trendability of a candidate condition indicator is defined as the smallest absolute correlation between measurements. For more information, see
In addition to using these functions at the command line, you can apply these feature-selection metrics in Diagnostic Feature Designer by selecting the prognostic ranking options.
Using the selected features to train an appropriate RUL estimation model is the next step in the algorithm-design process. For information, see Models for Predicting Remaining Useful Life.