Formulation of Model Stability Metrics for Remaining Useful Life Models of Engine Components

In the automotive industry, remaining useful life (RUL) predictions - an established prognostics approach replaces unplanned failures with informed preventive replacements. RUL models like any machine learning models are prone to degrade over time due to different reasons. For their stability, it is important to monitor and analyze their performance over time. The concept of model stability has been interpreted well in the context of RUL models. Higher values of model stability indicate higher confidence in the RUL model predictions while lower values point to the need for refinement and retuning of the RUL model. Such preemptive analysis ensures prompt corrective actions for component-model mismatch, thereby improving customer experience. This in turn benefits the manufacturer through additional business opportunities fostered by word of mouth and reputation. This paper discusses previous approaches and metrics to determine RUL model stability across different domains along with the novel ones. This includes both conventional approaches such as failure probability curve visualization with slope evaluation and newly proposed metrics such as the population stability index, coefficient of variation, asset fleet-wise metrics, and asset entity-based metrics. Stability metrics' values act as an indicator of the proper functioning of RUL models and hence would further contribute towards improving the RUL model. Our work shows specific applications for engine components namely, NOx sensor and turbo actuator. A comparative study of all approaches for determining the stability of RUL models, specifically, has been carried out through experimental analysis for better understanding. The work also recommends model stability metrics for the automotive industry with relevant analysis.