Within physical asset management, current centralized prognostics models face significant barriers due to data privacy concerns, while existing federated learning (FL) approaches often struggle with data heterogeneity across diverse assets. This study introduces FedCrossTeach, a novel decentralized framework that uniquely integrates a cross-teaching consensus mechanism with a multi-input multi-output (MIMO) architecture. Unlike traditional models, this approach enables the simultaneous estimation of remaining useful life (RUL) and fault classification in a privacy-preserving environment.
The methodology utilizes a dual-model structure comprising a private local long-short-term memory (LSTM) to capture asset-specific features and a shared Transformer for global knowledge aggregation. A novel cross-teaching protocol is implemented both within and between clients, where models exchange knowledge to build a fleet-wide consensus that acts as a regularizer against data noise.
Validated on C-MAPSS and N-CMAPSS datasets, the framework achieves a 25.2% improvement in root mean squared error (RMSE) compared to centralized benchmarks. Beyond statistical gains, FedCrossTeach allows maintenance managers to safely narrow uncertainty thresholds, extending asset life and reducing costs.
It offers a robust solution for cross-organizational collaboration, enabling competing entities to benefit from collective intelligence without the legal or commercial risks associated with raw data exchange.
