Towards a Framework for Multivariate Reliability Modeling: A Random Survival Forests Interpretability Case Study

As technical systems become increasingly complex, several variables must be monitored during field use to ensure technical reliability. This results in multivariate systems where failures cannot be attributed to a single variable, as interactions among variables can play a critical role. Moreover, real-world sensor data often includes units (candidates) that have not yet failed, introducing the challenge of censored data. This paper presents a literature review and conceptual study of machine learning (ML) methods applied to multivariate sensor datasets with and without failures, with a particular focus on the interpretability of Random Survival Forest (RSF) models. It examines existing ML approaches for identifying critical variables and determining their critical ranges, encompassing unsupervised, supervised, and survival modeling approaches. Additionally, it reviews how different modeling frameworks address censored data, feature interdependence, and degradation behavior. Building on this, the study explores RSF. It is highlighted as a promising example of a non-parametric method capable of capturing multivariate interdependencies. The paper discusses the possibility of applying RSF to time-to-failure data, assessing its potential to identify critical variables and determine the critical ranges that lead to failure. RSF is expected to serve as the basis for a general multivariate reliability model, providing interpretations into which parameters accelerate failure; an emerging, highly relevant yet underexplored area in industrial reliability research.The study is based on simulated multivariate timeseries sensor data representing light electric vehicle (LEV) fleet applications. These datasets serve as representative examples to illustrate the type of real-world information suitable for survival-based reliability modeling. These findings support proactive maintenance decisions. Overall, the study bridges predictive maintenance and explainable survival modeling with regard to technical complex systems, offering guidance on selecting suitable methods and enhancing understanding of time-to-failure prediction and critical variable identification.

Keywords: Explainable machine learning, Reliability modeling, Random Survival Forests (RSF), Multivariate time series, Automotive case study.