Reasons behind this project

Grid and industrial infrastructure operators make critical asset management decisions — when to maintain, refurbish or replace equipment — using increasingly large volumes of heterogeneous data: sensor time series, inspection images, maintenance logs, expert reports. Yet most reliability and survival analysis methods used today are fundamentally correlational: they identify statistical associations between variables but cannot distinguish cause from effect.

This creates a practical and costly problem. In the real world, maintenance is not assigned randomly: assets exposed to harsher conditions (corrosion, high load, pollution) tend to receive more maintenance, while also failing earlier. This confounding can produce deeply misleading conclusions — including the paradox that maintenance appears to shorten asset lifetime. Without a causal framework, operators cannot reliably evaluate the true impact of their decisions, nor simulate what would have happened under an alternative strategy.

Objectives

CSL-TD (Causal Survival Learning with Time-Dependent covariates) develops a rigorous methodological framework for causal survival analysis applied to industrial assets, enabling operators to quantify the true effect of controllable variables — maintenance strategies, operational decisions — on asset time-to-failure, using observational data alone.

The approach combines classical survival analysis (proportional hazards, accelerated failure time models) with causal inference techniques — causal graphs and potential outcomes reasoning — to explicitly represent confounding structures. A key originality is the integration of multimodal diagnostic data: time series, inspection images and textual reports are processed by a mixture-of-experts model to produce latent representations of asset degradation state over time, which then feed into the causal survival model as proxies for unobserved confounders. The framework enables counterfactual analyses (“what would have happened under a different maintenance policy?”) and feeds into an optimisation layer for determining optimal maintenance strategies. Methods are designed to be transferable across sectors — energy, rail, industrial processes.

Project partners

Hydro-Québec, RTE, CentraleSupélec and Université de Tarbes. The project is looking for additional partners.