Explainable AI Survival Models for Enhanced Reliability and Decision-Making in Industrial Operations

The growing complexity of industrial processes creates a need to have powerful predictive models in calculating the Remaining Useful Life (RUL) of assets that are considered as critical, which is the main task of the Predictive Maintenance (PdM). Survival Analysis (SA) models and, more specifically, the models that are rooted in Machine Learning (ML), like DeepSurv and Random Survival Forest (RSF), have shown to be more predictive than the traditional statistical techniques. However, complex ML models’ black-box nature is a serious impediment for implementation in a high-stakes industrial context, where trust, responsibility, and effectual insights are of primary concern. To overcome this challenge, the article used systematic reviews and applications of Explainable Artificial Intelligence (XAI) methods, namely SurvLIME and SurvSHAP(t) to MLbased SA models within the industrial operations context. A framework is offered that can be used to assess the practicability of explanations as time-dependent, which is consistent with the foundations of Explainable AI in Operational Research (XAIOR). On a representative industrial dataset, as this analysis reveals, DeepSurv has the highest predictive accuracy (C-index), but the combination of RSF and SurvSHAP(t) can give the most actionable and time-dependent information, which is essential in dynamic maintenance scheduling and developing operational trust. The research highlights the need of XAI to overcome the divide between developed ML forecasts and efficient and clear operational decision making.