Collaborative System Failure Prognostics via Federated Longitudinal–Survival Modeling

Research Question and Approach

Predictive maintenance in complex industrial systems relies on analyzing longitudinal sensor data to estimate remaining useful life (RUL) and failure risk. However, data privacy and proprietary constraints often prevent organizations from pooling sensor trajectories and failure records into a single central database. While federated learning (FL) offers a solution for decentralized training, standard survival models—specifically the Cox proportional hazards model—rely on global risk sets that couple observations across the entire dataset, making them incompatible with standard federated optimization protocols.

To address this, the authors propose a federated longitudinal–survival framework. They replace the continuous-time Cox model with a discrete-time hazard formulation. By discretizing the timeline into non-overlapping intervals, the authors transform the survival objective into a series of client-separable binary classification tasks. This allows each client to train a local temporal encoder (using LSTM and attention mechanisms) and a hazard head on their own data, communicating only model updates to a central server for aggregation.

The study evaluates the framework using the C-MAPSS turbofan engine degradation datasets, simulating decentralized environments with heterogeneous operating conditions. The results show that the federated model consistently outperforms isolated local training, where each site trains a model only on its own limited data. Furthermore, the federated approach achieves prognostic performance comparable to centralized training, demonstrating that the model effectively learns from distributed, diverse degradation patterns without compromising data privacy.

This research bridges the gap between advanced survival analysis and privacy-preserving machine learning. By enabling collaborative prognostics, the framework allows fleet operators and manufacturers to benefit from a broader range of failure experiences and degradation patterns, leading to more reliable RUL estimates. This is particularly valuable in safety-critical industries where data is siloed but the need for accurate, data-driven maintenance scheduling is universal.